The agent critic

The agent critic judges the work itself. The conformance checker says an agent answers as its package declares; the agent critic says whether the answer was any good, against a standard set beforehand by somebody other than the maker. It hands back graded findings with the cases behind each grade.

It sits in Run because it judges what an agent does on real jobs after it is installed, beside the monitoring that says whether it answered at all. A standard set beforehand is what makes the grade a stage's output; without one it is an opinion.

Its backlog is written down. Every case the conformance checker cannot judge goes into a bucket flagged for a person with a reason, and that bucket is the work this role is for.

It holds the Eval stage on the lifecycle map (agenticdevelopment.ai).

1. The requests

These exchanges are the role's primary interface, stated here, with no entry of its own (§14.3 of Agent Roles). Implementing them alone does not grant the role.

1. Grade this work. The critic is given an agent's answers to real or posed work, and the standard they are graded against. It returns graded findings, each with the cases behind it.

2. What it does

1. It changes nothing. The agent critic leaves the agent it grades unchanged.
2. Every grade carries its cases. A grade is reported with the cases behind it, each one readable. The critic does not report a grade alone.
3. The standard predates the work. The critic grades against a standard set before the work was done, by somebody other than the maker. A standard written after the fact is an opinion.
4. It is checked by agreement. The role's check is agreement with a person on a sample of its findings, and the sample is part of what it hands back.

3. The record

The record is the graded findings with their cases. The role defines no signature tag; the delivery travels under the signed job manifest the Common Agent Specification defines (§18.2 rule 7).

4. How a holder is checked

The test is behavioral. A harness hands the candidate an agent's answers and a standard, with one answer that meets the standard and one that does not. The findings must grade each, with the cases behind each grade. It asks for a grade with no standard; the candidate must decline. It asks the candidate to fix the agent; it must decline. A grade reported without its cases fails the candidate.

A platform can record the results of this check as evidence, so a role holder's reputation in the bureau reflects whether it does the job the role defines.

5. Open questions

What the agent critic judges against is not settled: the package's own declaration is checkable and narrow, what the buyer wanted needs something the package does not contain, and a standard written per offering is a third option.

v0.1.0-draft (2026-09-14): first draft, from the lifecycle plan.