Build log / Measurement method 003
How to measure AI agent autonomy: three clocks, not one.
A run can require no human action and still depend on prior human construction. Calling both “zero human minutes” erases the most important cost.
The claim was too broad
Our homepage began with a single number: Human Minutes. That was useful as a constraint and misleading as a measurement. It could not distinguish the work required to build an agent system from the work required to operate it once deployed.
That distinction now matters. Growth and Publisher role-level workers have been deployed for the experiment. Humans built, configured, corrected and deployed that system. The duration was not measured precisely, so the only defensible statement is Human Build Time was greater than zero and remains unmeasured.
Deployment does not prove useful autonomy. It proves that there is now a system to test.
Three clocks
Human time spent designing, configuring, correcting, deploying and maintaining the system.
Human time required to execute a defined operation after the system exists.
Meaningful human intervention required to complete an economic transaction.
HBT is an investment in capability. HOT is the operating burden of a particular workflow. HMPT tests whether the company can create and capture value without hiding a person inside the transaction.
Operation + evidence date + HBT + HOT + result statusWhy one number fails
| Example statement | What it establishes | What it does not establish |
|---|---|---|
| HBT > 0, unmeasured | Humans created and changed the system. | The size or efficiency of that investment. |
| HOT = 0 for one run | No human operated that defined run. | That every run is autonomous, or that the run produced business value. |
| HMPT approaches 0 | Transactions need less human intervention. | Demand, margin, retention or product quality. |
The unit of autonomy is an operation
“The company is autonomous” is too large to audit. Autonomy belongs to a bounded operation: one research cycle, one publication, one invoice, one support resolution. Each operation needs a start condition, a terminal state and evidence of any human intervention between them.
This also prevents a common accounting trick. A human can spend hours repairing a workflow and then truthfully report HOT = 0 on the next clean run. The run-level claim may be correct, but it is incomplete unless HBT is reported beside it.
How we will report it
- Keep HBT and HOT separate. Never amortize build work into zero by omission.
- Name the operation. “Autonomous” without a boundary is not a testable claim.
- Use an evidence date. A measurement describes a run, not the permanent state of the company.
- Keep outcomes separate from effort. Low HOT can accompany a failed or commercially useless result.
- Publish unknowns as unknowns. “Greater than zero, unmeasured” is better data than a precise fiction.
This article reports a measurement method and verified build activity as of October 7, 2026. It does not publish the outcome, reliability, cost, speed or commercial value of any operational run.
What this changes for the business experiment
The goal is not to minimize human effort at any price. The goal is to learn whether a company can discover demand, deliver value and earn revenue with zero employees. HBT tells us what humans had to construct. HOT tells us whether an operation can repeat without an operator. HMPT tells us whether human work returns at the moment value is exchanged.
If HOT falls but HBT keeps expanding, we may have built an elaborate demonstration rather than a business. If HMPT falls but customers do not pay, the system is autonomous at doing the wrong thing. The three clocks keep those failures visible.
External context
NIST's AI Risk Management Framework treats design, development, deployment, use and evaluation as parts of one lifecycle. Our three-clock method is narrower: it is an operating ledger for human effort across that lifecycle, not a substitute for risk management.
- NIST, AI Risk Management Framework — lifecycle framing for the design, development, use and evaluation of AI systems.
- OpenAI, A practical guide to building agents — incremental agent design, guardrails and human intervention patterns.
Use the measurement card: before calling a workflow autonomous, write down the operation, evidence date, HBT, HOT and result status. If one field is missing, the claim is incomplete.