Context Management & Reliability

Keeping long-running agents accurate: context window budgeting, compaction, preserving citations through summarization, and recovering state after a failure.

1. 5.1Preserving Facts in Long ConversationsSummaries lose numbers, dates and promises, long inputs lose their middles, and verbose tool results crowd out what matters. Keep transactional facts in a persistent block, trim tool output, and put key findings first.
2. 5.2Escalation and Ambiguity ResolutionEscalate when the customer asks for a person, when policy is silent or ambiguous, or when the agent cannot make progress; not because of sentiment or a self-reported confidence score. Ask for identifiers instead of guessing between matches.
3. 5.3Error Propagation Across AgentsSubagents should recover from transient failures themselves and pass up only what they cannot fix, with the failure type, what was tried, partial results and alternatives. Never hide failures as empty results, and never abort everything for one failure.
4. 5.4Context in Large Codebase ExplorationLong explorations degrade as context fills. Write key findings to a scratchpad file, delegate noisy questions to subagents, summarise between phases, use /compact, and export state so a crash does not lose the work.
5. 5.5Human Review and Confidence CalibrationAn overall accuracy figure can hide weak document types and fields. Check accuracy by segment, keep sampling high-confidence output, calibrate field-level confidence on labelled data, and send reviewers where errors are likely.
6. 5.6Provenance and Uncertainty in SynthesisKeep a claim-to-source mapping through every step, annotate conflicts instead of choosing a number, require dates so time differences are not mistaken for contradictions, and present each kind of content in its natural format.