Skip to content

Drill: runaway agent

Moduleops.10 · drill · ops, docs · Pass 11 · 1 to 2 h
You builddocs/runbooks/RunawayAgent.md
Testscourse/tests/ops.10 (named below, with why each exists)
Needsnone
Used byno code call site; this is an independent artifact
MilestoneMS-P11
  • Tie each claim to the exact platform evidence that Drill: runaway agent produces.
  • State the owner, success condition, and recovery action before declaring the work complete.
  • Record unresolved assumptions so the next review can test them.
Terminal window
ol start ops.10
ol tests ops.10
ol check ops.10

An agent loop can repeatedly call tools or consume its entire budget without producing useful work. The gateway and SDK need bounded behavior.

Set iteration, token, wall-clock, and tool concurrency budgets. Require gates for consequential actions, and make cancellation propagate through active calls.

Run a seeded agent that repeats a tool call and another that fans out parallel reads. Confirm the loop stops at the configured budget and reports a useful terminal reason.

TestKINDChecksWhy it matters downstream
test_artifact_paths_are_declaredartifactConfirms the submitted paths and required evidence exist, so a plausible narrative cannot pass without the reviewable deliverable.
#PitfallSymptomCaught by
1The canary advances before the rollback trigger is recordedThe exercise reports success without proving the old format or route can be restored.test_artifact_paths_are_declared
2The team watches request errors but omits the migration-specific signalA partial rollout can corrupt or strand work while the generic health check stays green.test_artifact_paths_are_declared
3No owner records the decision and verification evidenceThe next on-call cannot tell whether the incident is contained.test_artifact_paths_are_declared
DirectionModuleConnection
Backops.09Read this declared prerequisite before producing the artifact; use it to verify the relevant contract or evidence.
Backag.03Read this declared prerequisite before producing the artifact; use it to verify the relevant contract or evidence.
Backag.04Read this declared prerequisite before producing the artifact; use it to verify the relevant contract or evidence.
ResourceWhat to inspect
Incident drill catalogCompare this inject with the adjacent failure drills and use the same containment and verification fields.