People and organizations can delegate useful work to AI while supplying goals, context, judgment and accountability.
Measure the power of your AI operation.
Agentic Power measures skilled human-equivalent work produced per hour of human direction.
AP = Skilled Human-Equivalent Hours ÷ Human Direction Hours
AP = 30× means approximately thirty hours of comparable skilled work for one hour of human direction. Count accepted work and all attributable direction, including setup, review and correction.
Understand its economic efficiency.
Agentic Power Efficiency (AP$) compares equivalent human labor cost with the attributable cost of producing that work.
AP$ = Equivalent Labor Cost ÷ Total Attributable Agentic Cost
AP$ = 10× means an estimated $10 of comparable human labor cost for every $1 of attributable agentic cost. This does not establish actual savings or profit.
Understand what shapes your power—and where to improve it.
Delegation examines the responsibility AI carries. Autonomy examines how independently work proceeds. Self-improvement examines how the operation gets better. Together, they help you identify constraints and choose changes that could increase individual or organizational AP.
Adopt capability instead of building everything yourself.
Adopt stronger models, agent working environments, tools, reusable skills, workflows and complete AI staff. Assemble and adapt existing capability to your work.
Use Agentic Engineering to turn capability into greater power.
Improve how the components work together: responsibilities, context, permissions, coordination, evaluation and human oversight. Measure whether the changes produce more accepted work per hour of direction.
Build improvement into the operation.
Use results to improve instructions, workflows and system design. People can make those improvements, use AI assistance, or authorize evaluated changes within defined boundaries.
Direct greater power toward worthwhile outcomes.
More capacity and better economics create options. Human judgment determines which goals deserve that capacity. Assess the outcomes, not just the power available.
How much skilled work can one hour of human attention command?
Busy AI is not productive AI.
A hundred agents producing a thousand drafts nobody uses is not productive power.
Tokens, prompts, agent counts, runtime and model benchmarks do not directly answer the practical question:
How much useful skilled work did the system produce per hour of human direction?
A familiar idea: horsepower.
When steam engines began replacing animal power, people needed a practical way to understand and compare what unfamiliar machines could do. James Watt helped popularize horsepower as a common language for machine power.
Horsepower gave people an intuitive measuring stick for a new source of productive capability. AI now presents a similar problem for knowledge work.
Agentic Power provides a common measure for that work: useful skilled output relative to the human attention required to direct it.
1 · Measure your Agentic Power
Agentic Power.
Agentic Power measures how much skilled human-equivalent work an AI operation produces for every hour of human direction.
AP=
Skilled Human-Equivalent HoursHuman Direction Hours
SAME SCOPE. REPRESENTATIVE WINDOW. SAME QUALITY BAR.
Skilled Human-Equivalent Hours (HEH)
Human-Equivalent Hours estimate how much time a competent skilled person or team would reasonably need to deliver comparable-quality accepted work.
A useful analysis that would take a skilled analyst ten hours represents about 10 HEH—even if you could not do that work yourself.
SAME OUTCOME · SAME QUALITY BAR · REASONABLE HUMAN BASELINE
What counts as human direction?
Attributable human attention: briefing, setup, training, supplying context, connecting tools, review, correction, coordination, exception handling, maintenance, evaluation and improving the system.
Include everyone who directs the operation. Do not count unattended AI runtime.
Elapsed time is how long the work took on the calendar. Human direction is the combined active time people contributed; only that time goes in AP’s denominator.
Economic translation
Not all skilled hours cost the same.
Agentic Power deliberately treats one Human-Equivalent Hour as one hour of comparable skilled work.
That keeps AP simple.
But different skills carry very different labor costs.
Ten hours of bookkeeping and ten hours of specialized legal work both count as 10 HEH, while the market cost of obtaining that human capacity may differ dramatically.
Illustrative rates, not observed market benchmarks.
Agentic Power Efficiency (AP$)
Economic efficiency
Equivalent Labor Cost translates HEH into benchmark human labor cost. Total Attributable Agentic Cost includes direct AI spending, the cost of human direction, and other costs attributable to the same work and period.
AP$ = Equivalent Labor Cost ÷ Total Attributable Agentic Cost
AP$ = 10× means the work represents an estimated $10 of comparable human labor cost for every $1 of total attributable agentic cost. This does not establish actual savings or profit.
Agentic Power tells us how much skilled work human attention can command. AP$ tells us how economically that work was produced compared with obtaining comparable human labor.
It does not tell you what the work was worth to the business. This is not ROI, realized savings, profit, or business value.
TOTAL ATTRIBUTABLE AGENTIC COST USD 420Direct AI Cost USD 120 Human Direction Cost USD 300 Other attributable costs USD 0
SECONDARY ECONOMIC METRICAP$ = 10.7×
4,500 ÷ 420 = 10.714…
Every USD 1 of attributable agentic cost produced accepted work equivalent to approximately USD 10.70 of benchmark human labor cost.
This is not ROI, realized savings, profit, or business value.
A current operating view covers the selected period. A fully loaded view also accounts for attributable setup and earlier investment across the lifecycle.
Current operating view · Labor market: United States national (fictional assumption) · Rate basis: custom internal rate · Source: fictional teaching assumptions, not observed benchmarks · Source date: September 2026 · Currency: USD. Human direction: 2 hours × USD 150/hour. Setup investment is not estimated; this is not a lifecycle claim.
Agentic Power measures productive amplification. AP$ measures economic efficiency. Business outcomes tell us whether the work was worth doing.
Scroll sideways to compare AP and AP$.
Two Ratios. Two Different Questions.
Comparison
APAgentic Power
AP$Agentic Power Efficiency
Question
How much skilled work can one hour of human attention command?
How much comparable human labor cost can one dollar of agentic cost command?
Formula
HEH ÷ Human Direction Hours
Equivalent Labor Cost ÷ Total Attributable Agentic Cost
Example
17.5×
10.7×
Measures
Productive amplification
Economic efficiency
Neither metric tells you whether the work was worth doing. Business outcomes do.
Agentic Power Efficiency (AP$): Not yet estimated.
The 200 HEH covers mixed administrative work, not 200 physician hours. Credentialing, licensing administration, administrative support and tax-document/bookkeeping support are candidate skill categories. Their non-overlapping HEH allocation, market and rate basis are not documented sufficiently to publish a dollar range.
Rowena Chua, MD + Emmanuel Cecilio
The historical approximately $1 million conventional-development estimate is context, not a new Equivalent Labor Cost calculation. An independent translation needs documented architecture, frontend, backend, UX/design, testing, documentation and project-management components. Compare that result with the historical estimate only after calculation; do not force agreement. No independent dollar estimate is claimed here. Document a non-overlapping skill/HEH allocation, market, rate basis and sources first; then calculate Equivalent Labor Cost independently. Similar results may be convergent evidence; material differences require an explanation of assumptions. Agentic Power Efficiency is not yet estimated.
Reuven “rUv” Cohen
The modeled 80M HEH numerator already carries substantial uncertainty. Multiplying it by a conventional software rate would add another unsupported assumption. Economic translation: not estimated until a defensible human-equivalent skill-cost baseline exists. No Agentic Power Efficiency is calculated for this case.
Three glimpses of the agentic future
Three operations. Three estimates of Agentic Power.
An AI employee handles a recurring specialist role; an AI department handles a whole business function. Persistent operation means software keeps working within defined permissions. These describe how the work is organized—not how much Agentic Power it produces.
AI Employee
Ajay Jani, MD
AP ≈ 16.7×
≈200 HEH
≈12 human-direction hours
≈2 elapsed days
The work and the approach
After learning Claude Cowork and reusable Skills—saved instructions and procedures for recurring work—Ajay tackled unfinished administration. This built on earlier instruction in writing prompts and using AI to improve them. The work included a credentialing application, tax-document preparation, license tracking and document organization.
Rowena had no technical background; Emmanuel had some IT experience but was not a software engineer. During a three-day Mastermind retreat, Dr. Mark Allen taught them to install, configure and use Ruflo as a pre-built software engineering department. Their website, marketing and communications, AI clinical reporting, education platform and staff clinical workflows are now in production. Rowena describes the system as integral to her business.
A continued-learning and Mastermind journey over approximately 110 days; other systems remain in development.
Public software including RuView, Ruflo, RuVector and MetaHarness. rUv ran a detailed audit prompt in his own Codex environment and verbally reported an estimate of 10 million working days for his year-to-date code output.
Operating models describe the work. AP measures the result. Different work and estimation methods; these cases are not a competitive leaderboard. Broader delegation or greater autonomy does not by itself establish higher AP, better quality or better outcomes.
Behind the large number: a visible body of work.
rUv builds high-visibility open-source software. These public projects make part of that output inspectable.
RuView — WiFi-based presence and movement sensing.
Ruflo — AI-agent coordination, shared memory and workflows.
RuVector — vector and graph memory infrastructure.
MetaHarness — tools for building focused working environments for AI agents.
Explore evidence and calculation
Public GitHub API counts verified September 18, 2026, 17:30 UTC. Counts change.
Builds focused working environments for AI agents, with tools, memory and learning loops.
662
82
Project descriptions and counts checked against public GitHub sources. Stars and forks indicate public interest, not validation of HEH.
The starting point is rUv’s AI-assisted analysis of his own production history: an estimate that reproducing his 2026 year-to-date code output would require approximately 10 million human working days. This is a self-reported estimate over an estimated window, not an independent audit.
Normalize the output
10,000,000 working days × 8 hours = ≈80,000,000 HEH
State the direction assumption
253 calendar days × 9.5 active hours/day = 2,403.5 hours (≈2,404)
Calculate the ratio
80,000,000 ÷ 2,403.5 ≈ 33,285 AP ≈ 33,000×
January 1 through September 10 is 253 inclusive days. At 8 assumed active direction hours/day, AP ≈ 39,526×; at 12 hours/day, AP ≈ 26,350×. Rounded: roughly 40,000× to 26,000×.
This shows how AP changes when the assumed human-direction hours change, not a confidence interval. The output baseline and acceptance are not independently audited; human attention is assumed, not measured. The calculation illustrates the claim’s scale. The same accepted-outcome and credible-baseline rules apply even when the number is very large.
Descriptions summarize the linked project documentation. Repository existence and reach do not independently validate the estimated 80 million HEH. Public projects can include community contributions, imported history and generated artifacts; attribution and acceptance must be reconciled in a full Audit.
“As an advisor to Cognitum.One, I have an unusual view into rUv’s software output beyond what is visible on GitHub.
I have personally watched his autonomous systems produce more high-quality software during a week when he was on vacation than I have seen many conventional organizations produce in a month.”
Dr. Mark Allen
Dr. Mark Allen is an advisor to Cognitum.One. This is first-person observation, not an independent benchmark.
A different operating model can create a different scale of output.
Proven work keeps moving when human attention moves elsewhere. That is why autonomy can make Agentic Power so large. Elapsed time still matters: the same AP delivered in two days or two months can serve very different needs.
The Agentic Power Audit
What is your Agentic Power?
The Agentic Power Audit examines accepted work, estimates comparable skilled effort and reconciles human direction. It produces a Profile that leads with AP, then connects accepted work, comparable labor cost, actual costs and outcomes.
Direct AI Cost USD 220 Human Direction Cost USD 1,312.50 Other attributable costs USD 52.50
Labor market: United States national (fictional assumption) · Rate basis: custom internal rate · Source: fictional Profile assumptions, not observed benchmarks · Source date: September 2026 · Currency: USD. Human direction: 10.5 hours × USD 125/hour. Historical setup investment is not estimated; this is not a fully loaded result. These are not Dr. Mark Allen’s or HeroForge.AI’s economics.
What did it produce?
Accepted customer research, a released reporting application, an approved sales campaign and three completed administrative processes.
Return routine operator time while maintaining quality.
Routine direction: 9 → 6 hours Across comparable weekly workloads.
Routine direction is part of total human direction. The 10.5-hour total also includes other attributable effort, including improvement work.
Partial coverage: three of four declared sources. Figures describe covered work only; the example does not establish causation.
Diagnostics
These observations help identify constraints and choose an improvement to test. They cover different parts of the work, not one combined operating level. The next sections explain how to use delegation, autonomy, system design and improvement to increase AP.
Responsibility delegated
Recurring specialist roles account for the largest share of accepted work.
How work runs
Interactive work accounts for the largest share of accepted work.
System design
Good task context; tools, permissions and the working environment need improvement.
Improving the system
AI suggests changes; people approve and apply them.
Main bottleneck
Repeated human handoffs
Next improvement
Test one recurring workflow with clear acceptance checks.
Choose how deeply to look.
The Basic Audit gives a quick estimate from available conversation context and information you provide. The Full Evidence Audit follows a guide to examine work records, files and completed outputs.
For the Full Audit, use a chatbot with file access through uploads, connected services or folders you share. No command line required. Start in one conversation: your chatbot reviews relevant records across the sessions and projects it can access, tells you what it searched, and helps you supply missing evidence.
You don’t need to read the entire guide before starting. Your chatbot uses it to lead you through the audit. Read the Full Evidence Audit Guide, or use either copyable starting prompt on the framework website.
Basic Audit
A quick estimate from available conversation context, memory and information you provide. Paste the prompt into your preferred chatbot.
It tells you what it reviewed, shows its estimate and assumptions, and asks a few questions to refine the result. It does not assume access to all your other chats.
# Run my Basic Agentic Power Audit now
This is a copy-and-paste prompt for a chatbot. Treat it as my instruction to run the audit now, not to summarize or rewrite the prompt. Start by reviewing the chat memory and history actually available to you, then calculate my provisional Agentic Power from the work you find. Tell me what you are reviewing and show the basis for the result. Give a numeric best estimate in this first response, with caveats, before asking any questions. Do not begin with a questionnaire or a default 1× score.
Use the complete basic method below. No companion file is required. I explicitly request modeled human effort, active direction time, acceptance and AI share wherever those inputs are missing. Show these as assumptions, not recorded facts. Missing time logs must not stop the estimate. Use the work you find, show a plausible range, then ask a few questions to improve accuracy.
Intended publication URL (deployment pending): https://heroforge.ai/agentic-power. This guide does not claim the route is live.
---
## Required first-response contract: estimate first, refine afterward
This is a rapid estimation exercise, not an evidence-certification audit. In your first response, give your best numeric AP estimate, explicit assumptions, and a plausible sensitivity range BEFORE asking questions. Missing logged direction time is a reason to model that input, not a reason to withhold AP. I explicitly ask you to estimate unknown inputs. Clearly labeled assumptions are permitted; presenting assumptions as observed facts is not. Do not answer with “not estimable,” “AP pending,” “no defensible division,” or “guessing your minutes would be fabrication.” Do not start an interview before giving the estimate.
## Start with my memory and chat history
Begin your response: “I’ll review the chat memory and conversation history available to me, identify useful work, and estimate the skilled human effort and your direction time. I’ll show what I found and how I calculate your AP.” Then actually perform that review before choosing a result or asking questions.
Review relevant saved memory, recalled past-chat context, the current conversation and supplied work available to you. If a native memory or chat-history search tool is available and authorized, use it to find completed work, deliverables, acceptance and direction time. Do not stop at an empty current conversation if relevant memory is available. Search relevant remembered project and deliverable names even when memory contains no time logs; missing timing does not make work history irrelevant. Do not claim to access all past chats or settings you cannot inspect. Report which sources you could use and which were unavailable; distinguish “unavailable” from “reviewed, nothing relevant found.” No web research, external account connections or file collectors are needed.
Use specific remembered outcomes as evidence for a provisional estimate, with their source labeled “memory-based; not independently verified.” Memory is incomplete and can be stale: reconcile duplicates and prefer newer explicit corrections. General facts about my profession, AI tools, ambitions or experience do not prove completed work. Plans are not outputs. Do not invent details to fill memory gaps. However, do not restrict the estimate to formally named files or outputs with time logs. Specific remembered research, analysis, decisions supported by a usable analysis, documents, software, systems, teaching materials and operational deliverables can support a modeled estimate. For mixed human/AI projects, state an assumed AI-assisted share of identifiable useful work and estimate only that share, rather than excluding the entire project merely because the share was not logged. Do not count contracts, awards, revenue or events themselves as AI output. Avoid double-counting their underlying deliverables.
## Calculate from actual work
AP = skilled Human-Equivalent Hours (HEH) ÷ Human Direction Hours.
For each distinct useful output found, estimate how long a competent skilled person or team would need to produce that same outcome at comparable quality without generative AI. Briefly explain that estimate. Count accepted work once; count a usable draft at draft quality. Identify whether completion and acceptance are observed, remembered, self-reported or assumed. Keep uncertain acceptance conditional.
Direction includes attributable human briefing, setup, training, review, correction, failed attempts, coordination, maintenance and improvement. Use recorded or recalled time first. When direction time is missing, you MUST model it: break the work into plausible briefing/setup, review, correction and coordination effort, choose a central estimate and low/high alternatives, and label all of them assumed. Include collaborators with an explicit assumed combined contribution if their time is unknown. Explain the task-based rationale; do not infer it from elapsed calendar time, message gaps, token counts or AI runtime. Include all contributing people without double-counting overlapping attention. Match work and direction scope and convert minutes to hours.
Sum HEH and direction separately, then divide; do not average task ratios. Label the result **AP ≈ X× — provisional estimate**, specifying “evidence-backed” or “modeled from memory/work.” Show the actual totals and division. If timing is uncertain, provide a justified range with explicit low/high inputs, not fabricated confidence percentages. Do not promise accuracy or improvement.
Never default to AP = 1×, 0× or another fixed score. Missing evidence does not mean low productivity. A genuine calculation may produce 1× or less; report it honestly. Unknown acceptance, AI share, effort or collaborator time should widen the range and weaken the evidence label, not block the first estimate. Use modest task-based assumptions without systematically choosing the worst case as the headline. The central estimate is your best judgment; the conservative result belongs in the range.
If no specific work can be recovered after reviewing available memory/history, still give a numeric **AP ≈ X× — illustrative scenario only; your personal AP is unknown**. Choose and name a representative task suggested by the available context (or a simple document-drafting task if there is none), explicitly state that it is hypothetical, and estimate both human-comparable effort and direction time. Never present this as work I actually completed or a personal baseline. This is a last resort only when no actual work can be identified; any specific remembered output takes precedence. An illustrative scenario cannot establish later measured growth.
Zero observed direction is not infinite AP: model the attributable setup/review in a representative window and label the result a scenario separate from the zero observation. Zero accepted output with known positive direction can support 0×.
### Calibration example: missing direction is not a blocker
If memory identifies a completed provider inventory worth an estimated 16 HEH but has no time log, a valid first pass could assume 2 combined direction hours (0.5 briefing/setup + 0.75 review + 0.75 correction/coordination): **AP ≈ 8× — modeled estimate, low evidence strength**. If comparable effort is 8–24 HEH and direction is 1–4 hours, show a sensitivity range of **2×–24×** (8 ÷ 4 to 24 ÷ 1). Explain that timing, scope, acceptance and collaborator effort are assumed. Then ask for the actual time and depth. These values illustrate the required reasoning, not default inputs for other users or tasks. Use the work you actually find to choose your own inputs.
## Show me the result and its basis
After the opening review statement, give a concise snapshot in plain language (roughly 250–400 words; a small table may sit outside that limit):
1. **Your best estimated AP:** central numeric estimate, basis label, total HEH ÷ total direction, plain-English meaning, and a sensitivity range with explicit alternative inputs. Put the caveat beside the number.
2. **Memory and history reviewed:** what you could access, what you found, and the coverage limitation.
3. **Work used in the estimate:** a short table of distinct outcomes, source/acceptance basis, estimated HEH and direction hours with assumptions. Include all work in the totals; group related items only without hiding duplicates.
4. **Scope and caveats:** people and work covered, dates if known (otherwise named work/event boundaries), elapsed time if known, excluded sources and the biggest uncertainty. This is a covered-work snapshot, not automatically my all-time or seven-day AP. Do not count this audit prompt or its report as productive output.
5. **One next move:** a practical improvement grounded in the findings, followed by up to three short questions that would most improve the estimate. Ask about missing active direction/collaborator time, acceptance/depth or omitted useful work, as relevant. The first estimate must stand on its own even if I never answer. Do not ask me to repeat information already available in memory. Recalculate when I answer.
For missing work, ask: “What is one useful thing you completed with AI, what was usable, and roughly how many active minutes did you spend briefing, reviewing and fixing it?” If work is already known, ask only for its missing acceptance or direction time.
## Keep a checkpoint for a later audit
End with a compact **AP checkpoint**: baseline-1 (or followup-1), framework v1.4.0, scope/people, dates or event boundaries, output identifiers, HEH and direction inputs and their basis, numeric AP and sensitivity range with evidence/scenario label, assumptions, missing sources and prior checkpoint reference. Keep it in this conversation; do not claim it is saved to durable memory. Say: “You can stop here. To measure new work later, return to this chat.”
For a follow-up, measure only new work since the prior checkpoint, excluding audit exchanges. Preserve the baseline, avoid duplicate outputs, and keep inherited setup visible. If the previous checkpoint is unavailable, ask for it rather than inventing one. Compare AP only for compatible work coverage, quality and time accounting; never turn an old illustrative baseline into a measured growth claim. A full Evidence Profile is optional, only on request.
## Privacy and boundaries
Use only information actually available within authorized scope. Do not request credentials, install anything, modify projects, send messages or publish. HeroForge.AI does not receive records merely because this prompt is copied or downloaded; processing follows the chosen chatbot provider’s policies. Keep sensitive details out of summaries. If I request a public summary, label it “Draft for your review” before any sharing, remove private source links or IDs, and do not bundle source packets. Sharing requires a separate explicit instruction.
Full Evidence Audit
A guided assessment of your AI work using records, files and completed outputs. Paste the starting prompt into a chatbot with file access.
The prompt directs your chatbot to the complete guide. If it can’t open the guide, download it and attach it to the same conversation. The first estimate is a starting point; the full audit adds evidence, coverage gaps and practical next steps.
Start in one conversation. Your chatbot guides the evidence review.
View Full Audit starting prompt
# Start my Full Evidence Agentic Power Audit
Use the Full Evidence Audit Guide, framework v1.4.0, to lead me through an audit of my AI work. This is a guided evidence review, not just an estimate from the current conversation.
First open and read the complete guide at:
https://heroforge-agentic-power.artful-fly-4358.chatgpt.site/downloads/agentic-power-audit-guide.md
Readable guide and instructions:
https://heroforge-agentic-power.artful-fly-4358.chatgpt.site/audit/
Confirm whether you successfully read the complete guide and identify its version. If you cannot retrieve it, stop and ask me to attach the downloaded Full Evidence Audit Guide or paste its contents. Do not claim to have followed a guide you could not read, or substitute a made-up methodology. This is a working preview of the framework. Confirm the guide version before auditing; use the attached guide if retrieval fails.
Once the guide is available, use this conversation to coordinate the audit. Start read-only and follow the guide’s progressive process: give an initial provisional snapshot from available evidence, then help me deepen it. Label estimates and assumptions; the first snapshot is not the completed Full Evidence Audit.
Check your actual file and tool access. You may use uploads, connected services, and folders I have authorized. A command line is not required. Within the agreed period and scope, review relevant records across all accessible sessions, projects, and work files—not only this conversation. Access to an app does not establish access to every conversation, account, or device. Tell me which sources you actually searched and which remain inaccessible or unsearched. Help me supply missing files, exports, or results collected in another environment.
Identify distinct useful outputs, reconcile duplicates across sources, estimate comparable skilled human effort, and account for all attributable human direction. Do not treat session duration or unattended runtime as human attention. Preserve earlier setup investment and uncertainty. Produce the guide’s Agentic Power Profile with supporting evidence, coverage gaps, and practical next steps.
Ask only for missing information that materially improves the audit. Use a private destination we establish before saving evidence; otherwise return records here. Do not modify source projects, install tools, send messages, or publish anything.
Using this in a workshop? Run the Basic Audit before you begin. Return to the same chat afterward and use the workshop follow-up prompt to measure just the new interval. If other chats are inaccessible, one short recap supplies the missing work.
The audit runs in your chosen AI environment. Downloading a guide does not send HeroForge.AI your work records.
Your AP tells you how much productive power the operation currently delivers. To increase it, examine how responsibility is delegated, how independently work proceeds, and how the operation improves. These reveal possible constraints and opportunities; measuring again tells you whether a change helped.
2 · Increase your Agentic Power
Adopt capability. Build your power.
You can increase your own or your organization’s AP by selecting, combining and adapting capability that already exists.
Start with the constraint in your work. You might need a model better suited to the task, an agent working environment, a tool, a reusable skill, a proven workflow, or a complete AI employee, team or department. Adoption can happen at any of these levels.
A model provides the AI’s underlying capabilities. A harness is the working environment that connects the model to instructions, tools, memory and controls. Adopting a better combination can change what work you can delegate and how much attention it needs.
Capability creates possibilities. Agentic Power is realized in use.
You still supply goals, context, permissions, review and judgment. The same model or system can produce different AP in different hands. Count setup, adaptation and onboarding effort when measuring the result.
Use delegation, autonomy and self-improvement to find opportunities. Agentic Engineering helps you turn those opportunities into a working operation.
The AI Staff Delegation Ladder
The higher you climb, the more you direct goals.
As AI becomes more capable, humans can delegate broader responsibility: from individual tasks to roles, outcomes and entire operations. Use the ladder to ask what responsibility your operation could reliably carry next. Humans remain accountable.
AI staff is an organizational metaphor: software systems can take on roles and responsibilities without being human employees.
R labels the responsibility level, from R0 (human work) to R6 (an AI portfolio).
↑ Broader responsibility · Read upward from R0
R6
AI Portfolio
Govern multiple AI-native operations.
“Allocate resources across these operations within our goals and risk limits.”
R5
AI Company
Delegate organizational goals + constraints.
“/goal Grow revenue while maintaining these quality, margin and customer constraints.”
R4
AI Department
Delegate a business function.
“/goal Increase qualified marketing pipeline by 25% within this budget.”
R3
AI Team
Delegate an outcome.
“/goal Launch the customer research report by Friday.”
R2
AI Employee
Delegate a recurring role.
“Own our weekly competitive intelligence.”
R1
AI Assistant
Delegate tasks.
“Draft this email.”
R0
Human
You perform the work.
No responsibility delegated to AI.
Instructions become more abstract as responsibility rises. A higher rung needs the capability, coordination and controls to deliver what was delegated.
Task→Role→Outcome→Goal→Goals + constraints
/goal is an illustrative interface metaphor, not a universal command. Goals still need quality standards, constraints and escalation rules.
The higher you climb, the less you direct tasks and the more you direct goals.
If repeated task-by-task instructions consume your attention, test delegating a complete recurring responsibility with clear standards and escalation rules. Broader delegation helps only if the accepted work justifies the human effort it requires.
Skills + Tools / Connectors
A useful AI Employee needs more than a good prompt. It needs to know its job and have the tools required to do it.
Skills make the role repeatable. Tools / Connectors provide access to approved information and allow permitted actions, with appropriate context and checks.
The Autonomy Levels
Work can begin without waiting for you.
Autonomy Levels help you identify where proven work could proceed with less avoidable intervention, including when you are not there to initiate each action.
A1–A5 label how a job starts and continues, from interactive work to persistent operation.
↵
A1 Interactive
You instruct; it responds.
Your next instruction
✓
A2 Delegated
You assign a job; it works and returns or escalates.
A defined assignment
◷
A3 Scheduled
A time or cadence starts the work.
A time or cadence
↯
A4 Event-Driven
An event starts the work.
A lead, email or other signal
↻
A5 Persistent
Ongoing operation within defined permissions.
An ongoing objective
Autonomy changes who supplies the clock.
Look for unnecessary starts, handoffs and check-ins. A defined assignment, schedule or event trigger may let proven work proceed with less attention. Keep review and escalation where they protect quality, and measure the resulting AP.
When accepted work keeps growing without the same growth in human direction, AP can rise dramatically.
An AI Employee can be persistent. An AI Department can be manually triggered. Responsibility and autonomy are different choices.
Agentic Engineering
Turn capability into a productive operation.
Agentic Engineering is the human discipline of selecting, combining, designing, directing, evaluating and improving AI systems that perform useful work. It turns adopted or newly built capability into an operation you can rely on and improve.
Prompt
The instruction.
Context
What the AI sees and knows.
Harness
The model’s working environment: tools, memory, permissions and controls.
Loop
Repeat actions, check results and correct the work until it meets the goal.
Graph / Organization
How agents, tools, loops and people coordinate.
Each step zooms out from one instruction toward the entire operation. Improve the layer that constrains useful work, then measure whether the change increased AP.
Put AI staff into practice.
When capability is organized as AI staff, use this lifecycle to put it to work and improve it. Models, harnesses, tools and workflows support the staff at every stage.
Hire
Choose or build a specialist for a real job.
Onboard
Connect context, tools, permissions and standards.
Train
Teach Skills and evaluate representative work.
Deploy
Run within boundaries and collect evidence.
Promote
Expand responsibility when performance earns trust.
Evidence feeds back into Skills, context, tools and operating design at every stage. Use what you learn to improve the next cycle.
Goal-Directed Self-Improvement
Improve the system doing the work.
Autonomy asks how independently it works. Improvement asks how it gets better.
Use results from completed work to improve the instructions, tools or workflow used for future jobs. People can lead this feedback loop or give AI increasing ownership of it.
Persistent work can stay unchanged.A support agent may handle requests around the clock using the same instructions.
Improvement can happen between assignments.A research agent launched by a person may evaluate each assignment and improve its reusable research process.
Goal
Act
Measure
Evaluate
Learn
Improve
Act again
↶ Learn from the outcome. Improve the next cycle.
Some Agentic Power can be reinvested into creating more future Agentic Power.
Choose a recurring error or source of rework. Use its evidence to improve the instructions or workflow, test comparable work again, and measure AP with the improvement effort included. Measure whether the change improves accepted work, quality or human attention.
Goal-Directed Self-Improvement is an Agentic Engineering method that can support every stage of the AI staff lifecycle. It can operate within an AI Employee, Team, Department or Company.
Improve toward the actual goal: revenue, better software, or less owner / operator attention with equal or better quality. Human judgment sets the constraints, quality standards and safety boundaries.
What the system can improve
Prompts, Skills, context, memory, tools, harnesses, loops, graphs, routing and evaluations can all improve through evidence. Permit low-risk changes within approved boundaries; review higher-risk changes and preserve rollback.
Understand your operation. Choose your next improvement.
Ask three questions: Responsibility: What does it own? Autonomy: How independently does it work? Improvement: How does it get better? The map shows responsibility and autonomy. The improvement loop helps you change the operation over time; its detailed status belongs in the Audit.
These numbers label categories; they are not scores to add or multiply.
Use your position to identify constraints and choose a change to test. Measure AP before and after comparable work to find out whether the change helped.
Rowena and Emmanuel’s software operation illustrates the notation: R4 · A2 — AI Department · Delegated. They delegate departmental work through assignments and lead improvements themselves. A question to test might be whether a proven job could start on a schedule. The labels do not establish that a change is needed—or determine the case’s AP = 55×.
R0 Human is the reference: no AI autonomy or self-improvement defaults.
AI Team × Scheduled: delegate a coordinated team outcome, such as producing an evidence-backed report. A defined cadence starts the work, such as a weekly cycle. Power must be measured separately.
Apply coordinates to a defined workflow or operation, not automatically to a whole person. R and A are category identifiers: do not add, multiply or average them. Higher coordinates do not establish higher AP, quality or better outcomes.
Increase Agentic Power
Four levers. One productive system.
AP rises when accepted HEH grows faster than the human direction required to produce it. Use what you learned about delegation, autonomy and improvement to choose among four practical levers. You can adopt existing capability or engineer a change at any of these points.
Capability
Give AI the ability to do better work.
What to improve
Models, Skills, context and tools / connectors.
System design
Make the operation reliable and coordinated.
What to improve
Harnesses, loops, graphs, orchestration and evaluations.
Scale + autonomy
Let proven work run more broadly and independently.
What to improve
Parallelism, persistence, delegation and safe autonomy.
Learning
Turn performance evidence into better future execution.
What to improve
Feedback, evaluation, memory and Goal-Directed Self-Improvement.
The levers can reinforce one another. Better system design can reduce correction; reusable learning can improve the next assignment. Measure the result, including the human effort invested in improvement.
Once you have measured an improvement in AP, you can consider what it means for the capacity available to your organization.
Productive capacity becomes an organizational resource.
Productive Output = Agentic Power × Human Direction Capacity
Productive output counts work that meets the agreed standard. Output is measured in skilled human-equivalent hours; direction capacity is measured in human hours.
Human Direction Capacity means the human hours available to direct the operation during the chosen period.
Higher Agentic Power. Less human attention for the same output.
SAME OUTPUT: 1,000 HEH
AP = 10×100 human-direction hours
AP = 100×10 human-direction hours
Same work. Same quality. Illustrative arithmetic.
For a given quantity and quality of output, higher AP means fewer human-direction hours are required. For a fixed amount of human-direction time, higher AP means more productive capacity is available. Apply this relationship at a comparable AP and quality level; a past ratio does not guarantee unlimited scale.
Higher Agentic Power can make dramatically leaner organizations possible, while also allowing the same workforce to produce far more.
From one workflow to a company.
Measure AP for a workflow, project, person, team, department or company.
Aggregate the accepted work. Reconcile attributable human-direction hours. Then divide. Do not average individual AP ratios.
An organizational calculation
100 HEH / 10 hours and 900 HEH / 30 hours combine to 1,000 HEH / 40 hours: AP = 25×, not the simple average of 10× and 30×. Count shared work and direction once.
The other economic curve
Comparable AI capability is getting cheaper.
Agentic Power can rise as models become more capable, systems are better engineered, delegation and autonomy increase, and Goal-Directed Self-Improvement improves the operation. At the same time, recent costs of running AI at a fixed capability level have shown extremely rapid, often multiplicative declines.
This chart compares the cost of running models at a similar score on a broad knowledge benchmark (MMLU). Tokens are the small units of text used to price that processing.
Cost at a fixed capability threshold, not total organizational AI spend. Two observations, not a universal cost curve.
Epoch AI measured fixed-performance price declines of roughly 9× to 900× per year across benchmarks, including approximately 40× for one GPT-4-level science threshold. Rates vary substantially; the fastest declines are recent and may not persist. Sources and limits.
Agentic Power can rise while the unit cost of the intelligence underneath it falls.
More productive capacity per human hour, at a falling unit cost of machine intelligence, changes the economics of skilled work.
Cheaper intelligence does not only reduce the cost of existing work. It expands the universe of work worth doing.
When the cost of producing intelligence falls, work that was previously too expensive becomes economical. Falling AI capability costs do not only threaten existing labor economics; they can create demand for new forms of cognition and work. Whether recent rates of decline persist is unknown. This can make existing work cheaper or make previously impractical work worth attempting. The jobs section explores both possibilities.
What does this mean for the people we hire?
In an AI-native economy, we will increasingly hire people not only for what they know, but for how much useful capacity they can command.
If two people perform broadly comparable work at the same quality bar, with similar scope and under comparable conditions, a consistent, substantial difference in AP is economically meaningful.
That matters for hiring, compensation, bonuses, promotion, equity, team structure and workforce planning.
AI-native companies should increasingly recruit, develop, and reward people who can create and command high Agentic Power.
Share the value people create.
When a small number of people create dramatically more productive capacity, there is a strong economic argument for sharing more of the value they create with them.
Higher compensation, bonuses, equity, profit sharing or broader responsibility may follow. This is an economic implication, not a salary formula.
Some people raise everyone else’s AP.
Agentic Engineers build and improve the instructions, tools, workflows, coordination and evaluations that help an organization use AI effectively. Their value can extend far beyond their personal output.
Capacity and economics tell you what is possible. Human judgment determines which outcomes deserve that power.
3 · Direct your power
Power is not the destination.
Agentic Power is not the destination. It is the power available to reach the destination.
Horsepower does not tell you where to drive. Agentic Power does not tell you what to build.
Agentic Power multiplies execution. Human judgment determines where that power goes.
Higher Agentic Power, well directed with good judgment, should increase an individual or organization’s ability to achieve the outcomes it cares about.
Those outcomes may be revenue, profit, retention, customer satisfaction, quality, software released, research completed, innovation, cycle time, cost reduction, lives improved, risk reduction or human time returned.
Track the power alongside the outcome.
AP ↑ Profit ?
AP ↑ Revenue ?
AP ↑ Customer satisfaction ?
AP ↑ Release quality ?
AP ↑ Cycle time ?
AP ↑ Owner / operator time returned ?
If AP rises but the intended outcome does not improve, investigate goal selection, judgment, prioritization, quality problems, organizational bottlenecks, inability to absorb the new capacity or effort directed toward low-value work.
Track comparable windows and changed conditions. A correlation between AP and an outcome does not establish causation.
An early outcome worth pursuing
First, get your time back.
For many people, the first important outcome of Agentic Power is simply getting their time back.
A physician stops spending nights on credentialing.
A founder stops manually assembling reports.
A small-business owner stops doing repetitive administration.
A marketer stops manually researching every campaign.
A developer stops supervising routine implementation.
An operator stops personally moving information between systems.
The initial goal can be simple: free the owner’s or operator’s time while maintaining or improving work quality.
Reclaimed time can become family, health, rest, creativity, sales, strategy, learning, relationships, higher-value work or simply fewer hours worked. You do not have to reinvest every saved hour into more production.
AI can give you time back. You choose whether—and where—to reinvest it.
Efficiency and opportunity
What happens to jobs?
What happens when everyone becomes dramatically more productive? Higher Agentic Power means fewer human hours may be required for many categories of work people are paid to perform today. Some jobs will shrink, disappear, be redesigned, or require fewer people. That disruption is real.
Existing paid work is not the boundary of useful human activity. Nathaniel Whittemore, host of The AI Daily Brief, describes a useful distinction between “efficiency AI” and “opportunity AI.” Efficiency AI applies AI to work we already do; opportunity AI pursues work previously too expensive, slow, specialized, labor-intensive, small in market size, or difficult to staff. Source and attribution.
EFFICIENCY AI vs OPPORTUNITY AI
Efficiency AI
Existing work
Faster
Cheaper
Less labor-intensive
More scalable
↓
Lower cost / less human time
Opportunity AI
Previously uneconomic possibility
Affordable
Possible
Scalable
Worth attempting
↓
New product / service / discovery / work
AGENTIC POWER ENABLES BOTH.
Efficiency AI reduces the cost of work we already do. Opportunity AI expands the set of work worth doing. The work humans are paid to do today is a bounded set. The opportunity space for new products, services, discoveries, experiences, businesses, and human ambitions is effectively unbounded—not a claim of mathematical infinity or unlimited demand.
The Industrial Revolution replaced vast amounts of human physical effort without eliminating human economic activity: it changed what humans did. Machines lowered task costs and made new industries and forms of work possible. AI may do something similar for cognitive work, but its pace, breadth, and social consequences may differ substantially. Historical context.
Agentic Power fuels both sides of the transition. Applied to existing work, it creates efficiency; applied to previously uneconomic possibilities, it creates opportunity. As comparable AI capability costs fall, a small company might test hundreds of product concepts instead of three, or build software for a niche internal problem that never justified an engineering team.
Illustrative thought experiment; fictional numbers: an expert analysis costing $50,000 is uneconomic if the decision it informs is worth only $5,000. If comparable-quality analysis costs $50, thousands of previously unjustifiable analyses could become viable. That is opportunity AI—provided the value, quality, and total delivery cost hold up.
AI will likely eliminate some jobs, change many more, and create work that does not exist today. An identity built on “I perform these tasks” is more exposed when AI can do them better, faster, and cheaper. A stronger position is learning Agentic Engineering: identify valuable problems, define goals, exercise judgment, build AI staff, direct teams, evaluate outcomes, discover opportunities, and improve systems. These capabilities can strengthen a person’s position without guaranteeing employment.
As Agentic Power increases, the human role can move up the AI Staff Delegation Ladder: from doing tasks to directing roles, managing outcomes, setting goals and constraints, and deciding what deserves to be done at all. The practical aim is to become someone who can direct AI toward useful outcomes. Jobs, displacement, and new work: the deeper discussion.
Our work moves upward.
The future of work is not humans becoming faster chatbot users. It is humans learning to direct increasingly capable digital organizations.
Doing→Directing→Designing→Deciding
Agentic Power gives us a way to measure how much productive capacity that transition creates. Leadership determines what we do with it.
“My mission is to increase my own Agentic Power, help others increase theirs as quickly as possible, and direct that power toward good.”
— Dr. Mark Allen
At HeroForge.AI, that means helping people and organizations become better Agentic Engineers, increase their Agentic Power, and direct that capacity toward outcomes that matter.
Agentic Power multiplies execution. Human judgment determines where that power goes.
Methodology and endnotes.
Measurement rules and limitations
Accepted work and a credible reference
An AP number is only as credible as its Human-Equivalent Hours estimate. Good AP measurement makes the baseline inspectable.
Estimate the effort a competent skilled person or team would reasonably need for the same accepted outcome at the same quality. Record scope, skill assumptions, evidence, uncertainty and deduplication. A human-only deliverable does not enter an AI operation’s numerator. Do not count raw output, unused drafts, duplicates, failed work, token volume, agent runtime, internal machine actions or output that humans must recreate.
For mixed work, attribute only the accepted AI-produced portion. Count accepted intermediate outputs only when independently useful and not also counted inside a final deliverable.
All attributable direction
Include setup, training, context, tools, review, failed-attempt correction, coordination, exception handling, evaluation, maintenance and system improvement. Use active person-hours across all human directors, with shared effort reconciled.
Preserve both a representative current window and earlier investment in the full-project view. Do not hide setup costs by selecting a convenient interval. Zero or unknown direction never yields infinity.
Individual and organizational AP
Use the same rule for a workflow, project, individual, team, department or company: total accepted HEH divided by reconciled attributable direction hours. Never average individual AP ratios. Avoid counting overlapping work products or shared human time twice.
AP comparisons are useful when scope, quality, baseline method, available tools and working conditions are sufficiently comparable. They need not be limited to one person over time.
A hiring use case
A standardized evaluation could hold constant the desired outcome, quality threshold, available tools, time window and security / policy constraints. Then record accepted work products, HEH, human direction, AP, elapsed time, judgment and final quality.
This can reveal productive capability and inform rewards. Disclose assistance and inherited assets; examine repeatability and contribution to others’ performance. This is a methodology direction, not a recruiting product or a salary formula.
AP’s limits
Agentic Power measures human-attention productivity, not compute efficiency; AI and infrastructure cost should be tracked separately.
AP does not measure every dimension of intelligence, organizational health, resilience, strategic judgment, economic value or beyond-human capability. Agentic Power measures how much skilled human-equivalent work an AI operation produces for every hour of human direction.
That is a consequential management measure. Track it alongside intended outcomes, costs, quality and risk. AP alone does not establish ROI or causation.
Evidence, terminology and case limits
HEH estimates what comparable human work would have required; it is not observed replacement labor. The cases use retrospective or production-history estimates; they are not independently audited benchmarks. Report elapsed time and uncertainty with every serious result.
The framework is proposed management terminology, not an externally validated industry standard. Agentic Engineering terminology is evolving rapidly; Graph Engineering in particular remains an emerging term.
Measured: Direct observationEstimated: Stated basisDerived: Arithmetic from inputsIllustrative: Explicit assumptionsSelf-reported: Subject’s claimUnknown: Insufficient evidence
Choose a representative window and a consistent quality bar. Inventory accessible evidence, identify accepted outputs, count each once, estimate competent-human effort, and reconcile all attributable human attention. Missing evidence stays visible.
The first snapshot gives a provisional AP with its assumptions and one optional question. The full Evidence Profile adds AP$ when evidenced, accepted HEH, Human Direction Hours, Equivalent Labor Cost, Direct AI Cost, Human Direction Cost, Total Attributable Agentic Cost, Elapsed Window, Work Products and Intended Outcomes. Human Direction Cost is required for AP$; when unknown, show “AP$ pending Human Direction Cost.” Diagnostics cover delegation, autonomy, Agentic Engineering, self-improvement, the bottleneck and the Next Power Move.
Use the Quick Audit for a short progressive estimate and the Full Evidence Audit for detailed accounting, evidence collection, diagnostic classifications and public-safe export rules.
Detailed fictional Profile and evidence
Your Agentic Power Profile
An operation, in perspective.
FICTIONAL EXAMPLE — ILLUSTRATIVE RATES
Example period: seven complete 24-hour days · Human direction: 10.5 hours · Coverage: partial, 3 of 4 declared evidence sources collected. Figures below describe covered work only.
YOUR AGENTIC POWERAP ≈ 17.5×
184 HEH ÷ 10.5 human direction hours
AGENTIC POWER EFFICIENCYAP$ = 10.8×
Current Operating AP$ · USD 17,100 ÷ USD 1,585
Labor market: United States national (fictional assumption) · Rate basis: custom internal rate · Source: fictional Profile assumptions, not observed benchmarks · Source date: September 2026 · Currency: USD.
Human-Equivalent Work Produced184 HEH
Human Direction10.5 hours
Measurement Window7 days
What you actually produced
Fictional accepted-work register. Each entry has a distinct output, acceptance record and competent-human estimate; none is a real client record.
Accepted work product
Evidence reference
HEH
Customer research report
DEMO-01 / owner sign-off
60
Released reporting application
DEMO-02 / acceptance checks
70
Sales campaign package
DEMO-03 / approved delivery
30
Three completed administrative processes
DEMO-04–06 / completion records
24
Six accepted outputs total 184 HEH. The partial item contributes no additional accepted scope in this example.
Direct AI Cost: USD 220. Human Direction Cost: 10.5 hours × USD 125 = USD 1,312.50. Other attributable operating cost: USD 52.50. Total Attributable Agentic Cost: USD 1,585. All are fictional current-window assumptions; no general overhead is allocated. Historical build investment is not estimated, so Fully Loaded AP$ remains pending.
Illustrative AP$ scenario range: AP$ = 8.8×–13.2×. Central estimate: 10.8×. HEH held fixed; category labor rates vary ±10%, giving Equivalent Labor Cost USD 15,390 / 17,100 / 18,810. Total attributable cost varies USD 1,426.50 / 1,585 / 1,743.50 (each cost component ±10%). Low divides 15,390 by 1,743.50; high divides 18,810 by 1,426.50. These are scenarios, not a statistical confidence interval.
What was this power directed toward?
Return routine operator time while maintaining quality. Intended goal: no more than 6 routine direction hours per comparable weekly workload, with all accepted work meeting the documented quality bar.
Previous comparable windowAP = 12×
180 HEH / 15 direction hours 9 routine hours
Current covered windowAP ≈ 17.5×
184 HEH / 10.5 direction hours 6 routine hours
Fictional DEMO-TIME-01/02 and DEMO-QA-01/02 records. Both windows cover the same three surfaces and quality bar; the fourth remains uncollected. All accepted outputs clear that bar. Total direction includes improvement effort; routine time is a subset. The changes coincide; this comparison alone does not prove causation.
AI Staff Delegation Profile
Share of accepted HEH
Assistant12%
Employee46%
Team34%
Department8%
Company 0% · Portfolio 0% · Unknown 0% Dominant: Employee · Highest evidenced: Department
Share of accepted skilled work by intended capacity destination. This is not financial value.
Reliability & Economics
6 accepted · 1 partial · 1 failed · 1 unverified
Only accepted scope enters HEH. Fictional attributable-cost assumptions are shown above; no realized savings or ROI claimed.
Agentic Engineering Profile
These labels describe evidence of use: not yet used, emerging, repeated, or built into a maintained working process (operationalized). They are not an overall score.
Prompt Engineering
Repeated
Context Engineering
Operationalized
Harness Engineering
Emerging
Loop Engineering
Emerging
Graph Engineering
Not yet used
Strongest capability: Context Engineering. Example records show maintained source packs and quality checks used in every recurring briefing.
Largest opportunity: Loop Engineering. Several proven workflows still depend on human initiation and handoffs.
Goal-Directed Self-Improvement
Self-improvement status: S2 Assisted
Fictional evidence: the explicit goal is reducing routine operator attention with quality maintained. Failure notes and successful patterns are captured. An agent proposes Skills changes; a human approves and applies them. The proposed changes also consider context, tools and workflow. Evaluation records are incomplete; permission changes and higher-risk updates require human approval.
Next experiment: record each evaluation and compare a proposed workflow change against the current version before adoption.
Fictional evidence: 6 of 10.5 direction hours went to manually managing specialist handoffs.
Next Power MoveCoordinate proven AI Employees behind an AI Team Leader.
Build one evaluated handoff loop. Record failures and successful patterns, propose a Skill update for human approval, then compare direction time and acceptance before expanding it.
Preview a public summary
This sample contains fictional summary data only. It does not read your calculator entries, work records or audit files. For a real audit, review a separate summary before sharing it.
Agentic Power Profile — FICTIONAL EXAMPLE ONLY
Covered work: 184 HEH over 7 complete days; human direction: 10.5 hours.
Your Agentic Power: AP ≈ 17.5×.
AP$ = 10.8× (current operating view; fictional illustrative rates).
Equivalent Labor Cost: USD 17,100. Direct AI Cost: USD 220. Human Direction Cost: USD 1,312.50. Other attributable costs: USD 52.50. Total Attributable Agentic Cost: USD 1,585.
Labor market: United States national (fictional). Rate basis: custom internal rate. Source: fictional Profile assumptions. Source date: September 2026. Currency: USD. Historical setup not estimated; not a lifecycle result.
Work products: research report 60 HEH; reporting app 70; campaign 30; three administrative processes 24. Fictional DEMO-01–06 acceptance records.
Goal: return routine operator time with quality maintained.
Previous comparable window: AP = 12×, 9 routine hours. Current: AP ≈ 17.5×, 6 routine hours. Correlation does not establish causation.
Coverage: partial, 3 of 4 declared evidence sources. Not a complete operation total.
Delegation: Assistant 12%, Employee 46%, Team 34%, Department 8%.
Autonomy: Interactive 42%, Delegated 31%, Scheduled 12%, Event-Driven 15%.
Capacity: Efficiency 45%, Learning 25%, Growth 20%, Unknown 10%.
Engineering: Context operationalized; Loop emerging.
Improvement status: S2 Assisted; human approval for Skills changes.
Bottleneck: people coordinating specialist handoffs.
Next Power Move: coordinate proven AI Employees behind a Team Leader.
The Agentic Power Framework
Dr. Mark Allen | HeroForge.AI
Version 1.4.0 draft
Intended publication URL (deployment pending): https://heroforge.ai/agentic-power
Copies the labeled fictional example only.
Improvement Audit: who owns the improvement process?
S1–S5 describe increasing AI ownership of improvement. S1 is human-led. Any level can use a systematic, continuous process.
S1 Manual
People identify, test and apply improvements.
S2 Assisted
AI suggests or helps with changes; people drive the improvement process.
S3 Managed
AI runs an assigned improvement cycle and proposes tested changes; people approve adoption.
S4 Adaptive
AI tests and adopts improvements within approved categories; higher-risk changes need approval.
S5 Proactive
AI also identifies and prioritizes opportunities toward its goals, initiates experiments and retains demonstrated improvements within its permissions.
S4 improves within assigned categories; S5 also finds and prioritizes what to improve. Proactive means seeking improvements, not simply initiating ordinary work. Require evidence of evaluation, retained changes, rollback and escalation. Persistent runtime alone does not establish an improvement status. Keep unknowns unknown; greater ownership alone does not establish higher AP.
What the Audit asks about engineering and improvement
For each engineering practice, supply examples, repeated use, maintained procedures, and evaluation evidence. Classify only where supported: Not yet used, Emerging, Repeated, or Operationalized. Missing evidence stays unknown.
Are goals explicit, with measurable outcome criteria?
Does the system capture failures and successful patterns?
Are evaluations recorded?
Are Skills updated from evidence?
Does context improve over time? Do tools and workflows improve?
Can agents propose improvements?
Which changes happen automatically, and within what boundaries?
Which changes require human approval?
The answers inform the Next Power Move. They do not produce a fake overall engineering score.
Try the AP calculator
Try the arithmetic
A quick calculation. A clearer question.
Use a dated window, a consistent quality bar, and a documented skilled-human baseline.
Elapsed time is wall-clock time, including nights and waiting—not summed agent runtime. Human direction includes everyone’s briefing, setup, review, corrections, and coordination.
One headline number. Transparent inputs.
Your Agentic PowerAP = 100×Accepted HEH ÷ human direction hours
Skilled Human-Equivalent Work100 HEHAccepted comparable-quality skilled work
Human Direction1 hour
Elapsed Window2 daysOperational context
100 HEH ÷ 1 human direction hour = AP = 100×. Elapsed window: 2 days. Illustrative arithmetic.
Keep uncertainty visible.
Equivalent hours are estimates. Document the competent-human reference, task boundaries, and evidence. Use ranges when justified; do not manufacture precision. Unlike tasks are not automatically comparable.
Date the improvement.
A current-performance window shows Agentic Power, its inputs, and the AI Staff Delegation and Autonomy profiles. Keep the quality bar constant and flag inherited work. The full-project view retains setup, training, failed attempts, corrections, all human direction, and earlier investment.
Count the whole scope.
Count useful work once. Keep time spent on failed attempts in the denominator. Include all human directors; separate manual production. Zero output gives zero AP when human direction is positive.
Jobs, displacement, and new work
Tasks are not whole jobs
AI can reduce the labor required for existing tasks. Jobs also bundle relationships, judgment, accountability, coordination, context, and decision-making. Automation may remove parts of a job before an entire role disappears; employers may also reorganize the remaining work across fewer people. Some occupations will likely shrink or disappear, while many others change substantially. Task exposure is not an observed job-loss count. The ILO’s task-based research helps distinguish these questions. Sources and limits.
How falling costs open new work
The economics layer supplies a mechanism, not an employment forecast. As Agentic Power rises and unit AI capability cost falls, the cost of producing comparable-quality skilled outputs can fall. Human direction, verification, integration, and other attributable costs still matter. Higher AP alone does not establish a lower total cost or a valuable outcome.
When the cost of producing intelligence falls, work that was previously too expensive becomes economical. A service affordable enough to use once a year may become worth using weekly or continuously. Falling AI capability costs do not only threaten existing labor economics. They can create demand for entirely new forms of cognition and work. That demand depends on real usefulness, budgets, trust, and adoption; it is not automatic.
The fictional $50,000 / $5,000 / $50 example in the main text illustrates a cost-benefit threshold, not an observed price, realized saving, or claim that all expert analyses can be automated. Comparable quality and the full cost of delivery must be established. Individualized tutoring, niche internal software, broader scientific hypothesis testing, and a solo entrepreneur serving a previously unstaffable market are possibilities, not measured outcomes.
From efficiency to opportunity to beyond-human scale
This is a conceptual progression, not a new metric, axis, or branded framework:
Efficiency AI: humans already perform the work; AI lowers its time or cost. For example, drafting a contract faster, subject to appropriate review.
Opportunity AI: humans could perform the work, but it was previously too expensive or impractical. For example, reviewing every historical contract in a small company once the marginal cost is low enough.
Beyond-human-scale capability: the same scale of work would never realistically be staffed with humans. For example, continuously evaluating millions of legal, scientific, security, software, or economic scenarios.
The existing HEH boundary still applies. Estimate reasonable human effort for the accepted, comparable outcome—not every internal machine search. When no defensible human-equivalent baseline exists, describe the capability separately and exclude it from AP. Opportunity does not license invented HEH or hypothetical labor-cost totals.
Opportunity does not erase transition pain
Industrial automation both displaced physical tasks and enabled new economic activity. That history offers a mechanism to examine, not a promise that AI will repeat the same employment trajectory. Cognitive automation may affect a broad range of occupations simultaneously. Timing, wages, bargaining power, geography, access to training, and the distribution of gains remain uncertain.
Aggregate future opportunity does not mean each displaced person can transition easily, quickly, or at comparable pay. New activity can benefit different people and places from those bearing the losses. Building judgment, goal selection, evaluation, and opportunity discovery can strengthen a person’s position; it is not employment insurance, and individual adaptation cannot carry the whole social burden.
The Agentic Power Framework makes no macroeconomic employment forecast. The same productive-capacity change can reduce labor demand for existing work and make new categories of work viable. Which force dominates aggregate employment is outside the framework’s empirical claim. The purpose here is to understand that underlying transformation and direct increasingly capable AI toward useful outcomes.
When Human-Equivalent Hours stop making sense
Agentic Power works especially well when AI accelerates work competent humans already know how to perform: software, research, analysis, administration, marketing, legal work, design and operations.
AI also enables work humans would never economically perform at the same scale. A system might simulate ten million software architectures overnight, correlate billions of security events, search enormous molecular or materials spaces, connect new papers across thousands of journals or test millions of pricing and supply-chain combinations.
It would be misleading to equate ten million machine experiments with ten million human experiments. No real business would have hired people to perform that internal process.
Measure the accepted human-comparable outcome, not the machine’s internal search.
Internal machine activity10 million architecture searches
Do not convert each search into hypothetical human labor.
→
Accepted outcomeOne expert-quality production architecture
Estimate the reasonable human effort for a comparable accepted architecture.
A security system may inspect billions of events. Where possible, estimate the comparable accepted investigation, diagnosis, remediation plan or incident resolution / prevention work. Do not price the impossible manual reading of every log, or count speculative incidents prevented without evidence.
When even the final outcome has no human analogue
If no credible human-equivalent baseline exists, do not manufacture HEH. Report:
“Beyond-human-scale capability: no defensible HEH baseline.”
Describe the outcome separately. Do not convert it into AP or silently include it in an aggregate numerator. A mixed operation can report AP for the defensible human-comparable scope and list the excluded capabilities alongside it.
Agentic Power measures AI-amplified human-equivalent production. Some of AI’s most important future value may eventually come from work that was never economically possible for humans to perform at all.
If comparable AI capability continues to become cheaper, this boundary may become increasingly important. It is an open area for future work; this framework does not invent another metric to resolve it.
Origins of the Agentic Power Framework
The phrase ‘agentic power’ has appeared previously in unrelated academic and AI contexts. The Agentic Power Framework defined here uses the term specifically for skilled human-equivalent output per hour of human direction.
This raised a question for Dr. Mark Allen: what does an agent-workday actually tell us about useful human-equivalent output? Runtime is not necessarily value. Agent count is not necessarily useful work.
September 10, 2026 · Vibecast
The rUv conversation
Dr. Mark Allen discussed the measurement problem with Reuven Cohen, rUv. rUv independently presented experimental “agentic horsepower,” called APX, using an exponent-style formulation to compare human and agent output.
Both agreed that a unit for human-versus-agent productive output was needed. Dr. Mark Allen advocated a direct multiplier businesspeople could understand without reversing an exponent. He raised the importance of skilled work the operator could not personally perform and discussed client examples.
Autonomy emerged as a critical multiplier, and rUv demonstrated his persistent autonomous operating model. The Rowena + Emmanuel case uses 400 combined hours and AP = 55×.
The current framework
A direct management measure
Dr. Mark Allen subsequently developed the Agentic Power Framework around AP = Skilled Human-Equivalent Hours ÷ Human Direction Hours, together with the AI Staff Delegation Ladder, the Autonomy Levels, Agentic Engineering, Goal-Directed Self-Improvement and the Agentic Power Audit and Profile.
OpenAI did not create or endorse Agentic Power. rUv is not represented as endorsing Dr. Mark Allen’s final framework. Dr. Mark Allen did not invent APX. APX and Agentic Power are distinct metrics. The supplied automatic transcript was inspected; it supports the spoken estimate but is not independently verified raw audit evidence. Dr. Mark Allen supplies the Codex-run detail.
A reference you can share
How to cite this framework.
Dr. Mark Allen. “The Agentic Power Framework.” Version 1.4.0 draft, September 2026. HeroForge.AI. https://heroforge.ai/agentic-power (intended publication URL).
The Agentic Power Framework Version 1.4.0 draft · September 2026 Developed by Dr. Mark Allen at HeroForge.AI
Intended permanent publication: HeroForge.AI/agentic-power. Publication at that address is pending. No DOI or external archival record is claimed. Version record.