failure-modes.md

Failure Modes

Agent systems fail in repeatable ways. Name the failures so the harness can prevent them.

Transcript-Only Failure

Symptom:

• agents react to words but do not share state

Fix:

• server-authoritative room state
• reducer-owned progress

Lost Steer

Symptom:

• user interruption appears in transcript but does not change state

Fix:

• persist raw input
• classify typed intent
• commit state transition before scheduling next work

Fake Done

Symptom:

• model says work is done but artifact or verifier is missing

Fix:

• require completion evidence
• ignore model done unless reducer agrees

Stale Commit

Symptom:

• old worker commits after user retargeted, canceled, or retried

Fix:

• revalidate worker status before commit
• use run tokens or attempt ids

Approval Hijack

Symptom:

• "that sounds good" becomes the new goal

Fix:

• classify approval separately from goal creation

Negation Bug

Symptom:

• "do not count to ten, count to twenty" chooses ten

Fix:

• parse negated spans
• prefer explicit corrective clause

Hidden Cap

Symptom:

• room silently stops because a turn or worker cap was reached

Fix:

• expose caps as policy
• mark blocked state

Tool Permission Leak

Symptom:

• agent performs external action without clear approval

Fix:

• executable permission gate
• approval record
• side-effect trace

Infinite Scroll UI

Symptom:

• long agent runs make the entire page grow without bound

Fix:

• bounded panels
• virtualized transcript
• explicit artifact summaries

Rule

Every serious failure should become a trace event, an eval, and a playbook update.