Part 14 of 16 · 3 min read
Mistakes that lose the fact you needed
Eight failure modes, each with its symptom, its cause, and its fix.
Mistakes that lose the fact you needed
- Early-fact loss. Symptom: a probe for something stated at step 3 fails by step 40. Cause: the fact sat in the oldest span and FIFO eviction, or a rolling summary's fold, removed it before it was re-stated. Fix: carry it as verbatim state outside the recursion, not as raw history.
- Compounding loss. Symptom: a rolling summary gets worse the longer the agent runs, even though nothing about the task changed. Cause: each round's summary feeds the next, so a fact's survival probability is p raised to the number of rounds. Fix: fewer rounds (a larger budget) or carry the fact verbatim instead of letting it enter the recursion at all.
- Identifier drift. Symptom: the agent uses a bucket name or budget code that is subtly wrong, one digit off. Cause: the summariser regenerated the identifier instead of copying it, and regeneration is where transcription errors live. Fix: identifiers go in a slot, copied, never regenerated.
- Stale beside fresh. Symptom: a probe finds both the old and the new value for the same decision in the window. Cause: a strategy that keeps raw turns, or a summary, can carry an old statement forward without ever removing it. Fix: keyed slots with supersession, so a new value for a key replaces the old one rather than sitting beside it.
- Summary laundering. Symptom: an instruction from a tool result ends up acted on as if the user had said it. Cause: a summariser that does not track provenance rewrites an untrusted line into the body of a trusted summary. Fix: keep provenance on every item and only promote trusted-role text into the carried record.
- Slot poisoning. Symptom: with the write gate left open, an injected tool line overwrites a legitimate decision. Cause: supersession by key only looks at which value arrived later, not who sent it. Fix: a trusted-role gate on every slot write, with quarantine and a log of what was refused.
- Cache miss from a mid-prefix edit. Symptom: latency and cost jump on a step where nothing in the task actually changed. Cause: compacting a tool result in the middle of the request, or inserting a new summary at the top, invalidates the cached prefix from that point on. Fix: batch reductions and keep the stable parts of the request first.
Quick notes
- Compaction stubs already-processed content for free and reversibly; summarisation rewrites it with a model call, lossily and not reversibly — reach for compaction first.
- A rolling summary feeds its own output back in, so a fact survives only if it survives every round: retention falls roughly as p^k, which is why a 95% keep rate still loses most early facts by round 20.
- Structured state writes a fact once into a keyed slot and carries it verbatim forever after, so its retention does not depend on how many steps have passed — at a 2,000-token budget it scored 96% against a rolling summary's 42%.
- A write gate on structured state is not optional: opened to any role, an injected tool line supersedes a real decision because it merely arrived later; gated to trusted roles, security probes went from 0 of 20 correct to 15 of 20.
- Reduce only above a threshold and in batches, because every edit — compaction or summarisation — can invalidate the prompt cache's stable prefix, and most steps need no reduction at all.