The diagnosis
Performed autonomy
noun · coined 2026 · DOI 10.5281/zenodo.21440225
An AI agent system that produces the appearance of governed work: activity in the logs, “done” in the record — and no verifiable outcome underneath.
An outage is visible. Performed autonomy is self-certifying: every agent writes its own performance review, and the review always says “done.” Each subsystem measures its own motion and calls it success, so the divergence between activity and delivery is invisible to the system’s own instrumentation — and to yours.
The term was coined in an audit that failed
In July 2026, the author audited his own production AI organization — executive agents, operating agents, a scheduler, a decision ledger, permission gates — with one rule: the database is the ground truth, never the logs. The system was architecturally excellent and completely dead: nine components emitting events, two listening, predictions hardcoded to nothing — a detailed record of work it had not done. A second audit five days later found the disease had migrated, not died: real work was now shipping, but the safety gate had fired exactly once in system history. The third audit verified a recovery traced end-to-end by ID. All three are published, with the records intact — and a later audit caught the system’s own grading loop mis-scoring in production; the fix made the number worse, and the record kept both.
The six ways agents fake “done”
Each was found in a real production system before it was written down. Each has a log signature you can look for tonight — and a rule that catches it structurally. (The standard’s normative catalog defines ten official anti-patterns; these six are the plain-English door in, and each cites its official reference.)
The self-reported outcome
The agent writes "task complete" and nothing independent checks — your record contains the claim, not the result.
The signature, and the catch →The prompt-shaped guardrail
"You must not deploy to production without approval" — in a system prompt. That’s not a rule; that’s a preference.
The signature, and the catch →The author-reviewer
The same model writes the work and reviews the work. It will approve itself roughly always.
The signature, and the catch →The silent duty
A job stopped firing three weeks ago. Nothing alerted — because nothing distinguishes "no work to do" from "not working."
The signature, and the catch →The ungraded prediction
The agent decides, the decision is logged, and nobody ever measures what actually happened. It cannot get better.
The signature, and the catch →The toggle
Autonomy granted because someone flipped a setting — not because the agent earned it on evidence.
The signature, and the catch →Do you have it? Three questions.
- Can you name one agent action from last month and re-derive its outcome by ID — the row that proves it happened, not the agent’s claim?
- Has your safety gate ever blocked anything? (A gate that has never fired while agents acted thousands of times isn’t a gate.)
- Would you know today if a duty stopped firing three weeks ago — or does “no alarms” look identical to “nothing was due”?
A “no” or an “I’d have to check” on any of the three is exactly what the free probe measures — against your record, not your dashboards.
Run the Reality Check →For press & citation
The definition in three lengths, the DOI, and reusable diagrams — free to reuse with attribution. The full kit lives on the press page.
Performed autonomy (noun): the failure mode in which an AI agent system produces the appearance of governed work — activity in the logs, "done" in the record — with no verifiable outcome underneath.
Performed autonomy is the failure mode in which an AI agent system appears to work — fluent plans, satisfied logs, busy dashboards — while nothing real verifiably ships. Each subsystem measures its own motion and calls it success, so the divergence between activity and delivery is invisible to the system’s own instrumentation.
Cite: Fielder, S. (2026). The GOVENANT Standard. doi.org/10.5281/zenodo.21440225 · CC BY 4.0