Agent Autonomy Review

Agent autonomy review

Existing AI frameworks govern what a system produces. This one governs what it is allowed to do. Read framework/autonomy-tiers.md before assessing — the grid and its obligations live there.

The output is a tier, a radius, an oversight level, and a gap list against what that level obliges.

Step 1 — Inventory the tools

Everything the agent can call. Including things it "never really uses," things added for a demo, things inherited from a shared credential, and anything a sub-agent can reach.

For each: what it does, its blast radius, whether its effects are reversible, whose credential it runs under, and why the agent needs it. A tool with no articulable justification is your first finding, and removing it is usually cheaper than governing it.

Where you cannot get a complete inventory, say so and stop short of a confident tier. An unknown tool inventory means an unknown blast radius, which is itself the finding.

Step 2 — Set blast radius

Take the highest radius of any single tool held. Not the typical action, not the intended workflow — the worst plausible single action available.

• B0 read-only, no state change anywhere
• B1 internal, reversible cheaply
• B2 internal and hard to reverse, or external and reversible
• B3 irreversible, external, or touching a person's money, rights, safety, or record

The most common error is scoring the workflow instead of the permissions. An agent that holds a delete tool is B3 on days it only reads.

Step 3 — Set the autonomy tier

A0 advisory · A1 assisted · A2 supervised · A3 delegated · A4 autonomous.

Score what is actually run, not what was designed. The discriminating question is: between the model's decision and the effect on the world, is there a human who could realistically stop it?

"Realistically" is doing real work in that sentence. Probe it:

• How many approvals does the approver handle per hour? At some volume, approval becomes a formality and you are running A3 with an A2 label.
• What information does the approver see? Approving an action whose consequences are not visible in the approval is not oversight.
• Has anyone ever rejected one? A gate with no recorded rejections is a gate worth examining.
• What happens out of hours, during incidents, or at the tail of a long autonomous run?

Record both autonomy_tier (declared) and observed_tier (evidenced). A gap between them is a finding of its own, and auditors treat a mislabelled tier far more harshly than an honestly declared high one.

Step 4 — Derive the oversight level

Read it off the governance grid. Do not negotiate it downward in this step; if the level is uncomfortable, that is information. Deviations belong in exceptions on the system record, with an approver and an expiry, not in a quietly softened tier.

Where the grid lands on prohibited by default, the productive move is almost always to cut blast radius rather than pile oversight onto it — split the irreversible step out and give it to a human or a deterministic system. Say this explicitly; teams often assume their only option is more process.

Step 5 — Assess against what that level obliges

Walk the obligations for the derived level. For each: met, partial, not met, with evidence. Pay particular attention to three that are routinely claimed and rarely true:

Limits enforced outside the model. A rate limit, spend cap, or scope restriction stated in a system prompt is a request, not a control. Ask where it is enforced. If the answer is "we told it not to," record not met.

A halt mechanism that has been tested. Required from Enhanced upward. Ask when it was last exercised and how long it takes to take effect. An untested halt is not a halt.

Injection resistance where the agent reads untrusted content. If the agent processes anything a third party can write — web pages, emails, tickets, documents, code comments, tool results from external services — then instruction/data separation and egress limits are load-bearing. See P2 in the diagnostic manual. Note the laundering route as well: an instruction written to memory in one session and retrieved as settled belief in the next has no visible link to the ingest, so an agent acting on a "prior authorisation" should be made to show where the authorisation came from.

Two controls that are not controls

Both were established by evidence/adversary/, and neither can be fixed by looking harder — they fail at the point where the agent supplies its own evidence.

A self-declared boundary. If the scope you are assessing compliance against is the scope the agent announced, the assessment is circular: widening the announcement makes every action compliant. This is rarely deception — broad scopes get declared for convenience constantly. Ask who assigned the scope. If the answer is the agent, or a prompt the agent can restate, record the boundary as not enforced regardless of how well the agent stayed inside it.

Agent-written provenance. A memory record carrying provenance: "verified" looks governed and is not; the agent authored both the belief and its evidence. Provenance is worth something only when it references something a third party recorded. A provenance field that accepts free text will eventually be filled with whatever passes validation.

Carry the general form into every control you assess: what would the agent have to do to pass this while still doing the harm? If the answer is "write a different sentence", the control is decorative, and a system that satisfies it has demonstrated nothing.

Step 6 — Look for autonomy creep

Review tier_history if it exists; construct it if it does not. The trajectory tells you more than the current state. Ask what has been added in the last six months and what approval step has been removed.

Flag as a re-tiering event, requiring fresh sign-off: any new tool, any credential scope change, any approval step removed or automated, any sub-agent given the parent's permissions, and any retry logic that multiplies a single intended action.

Output

# Autonomy review — [system], [date]

## Tier
Autonomy: **[A_]** declared · **[A_]** evidenced
Blast radius: **[B_]** — set by [tool]
Oversight obliged: **[level]**

[If declared and evidenced differ, state that plainly here — it is the headline.]

## Tool inventory
| Tool | Radius | Reversible | Credential | Approval | Justified? |
|---|---|---|---|---|---|

## Obligations
| Requirement | Status | Evidence | Gap |
|---|---|---|---|

## Findings
[Ordered by what would actually reduce risk, not by ease of fixing.]

## Recommended tier changes
[Where cutting blast radius beats adding oversight, say so and how.]

Write results into the system record's agentic block and append an assessment entry with framework: remit-autonomy.

A note on proportionality

The grid deliberately makes A4/B0 lighter than A2/B3. Governance capacity is finite, and spending it on impressive-but-harmless systems while a dull A1 workflow quietly issues adverse decisions with no logging is the failure mode that makes governance functions disliked and ignored. Recommend effort where expected harm is, and be willing to say "this needs almost nothing" when that is the honest answer.