A persistent structured object holding transactional facts — amounts, dates, order numbers, statuses — injected verbatim into every prompt. Summarization turns "$47.20" into "large amount"; keeping it outside history prevents precision loss.

Models reliably process content at the start and end of long inputs but may skip material buried in the middle. Counteract by placing key findings first and organizing detailed results with explicit section headers.

A single order lookup can return 40+ fields when only 5 matter. Untrimmed outputs accumulate tokens disproportionately to their relevance, degrading context efficiency across turns.

Extract and persist structured issue data (order IDs, amounts, statuses) into a separate context layer for each issue — preventing distinct cases from blurring together across turns.

Structured data — key facts, citations, relevance scores — rather than verbose content and reasoning chains. This includes metadata like dates and source locations for accurate downstream synthesis.

(1) Customer explicitly requests a human. (2) Policy is silent or ambiguous on the specific request. (3) Agent cannot make meaningful progress. Sentiment and confidence scores are not triggers.

Sentiment measures emotional tone, not case complexity. LLM self-reported confidence is poorly calibrated — the agent is already incorrectly confident on the hard cases it escalates incorrectly.

Route to a human agent immediately — no investigation, no attempt to resolve. Honoring an explicit human request without delay is a hard rule encoded in the system prompt.

Acknowledge frustration and offer resolution — the issue is within the agent's capability. Escalate only if the customer reiterates their preference for a human agent.

Ask for an additional identifier — email, order number, or ZIP code — to disambiguate. Never select a customer record by heuristic when multiple matches exist.

failure_type · attempted_query (the exact query tried) · partial_results (what was found before failure) · alternatives (other approaches to try). The coordinator needs all four to recover intelligently.

Access failure: the query couldn't run (timeout, permission denied) — a retry decision is needed. Valid empty result: the query ran successfully but matched nothing — no retry needed, the absence is the answer.

It hides the failure type, the attempted query, and partial results. The coordinator cannot make an intelligent recovery decision — retry, reroute, proceed partial, or escalate — without this context.

Handle locally only transient failures the subagent can resolve (brief timeout — retry once). Propagate errors the subagent cannot resolve, always including what was attempted and any partial results.

After many turns, the model starts giving inconsistent answers and citing "typical patterns" rather than specific classes or structures it actually discovered earlier in the session.

A file (e.g., findings.md) on disk where agents write key findings as they discover them. Re-read before subsequent questions to counteract context degradation and prevent drift back to "typical patterns."

Raw file listings and dependency traces stay in the subagent's isolated context. The coordinator receives only a compact summary — its context stays uncluttered for high-level coordination.

Reduces context usage when an extended exploration session has filled the context window with verbose discovery output. Allows the session to continue without hitting context limits.

Each agent exports state to a known location. On resume, the coordinator loads a structured manifest and injects agent state into prompts — the exploration resumes without restarting from scratch.

A 97% overall metric can hide one document type or field failing at 71%. Always analyze accuracy by document type and by field before reducing human review on any segment.

Sampling high-confidence auto-accepted extractions by document type and field (not uniformly) to measure ongoing error rates and detect novel error patterns before they propagate across large volumes.

Raw model confidence is not calibrated out of the box. Calibration against a labeled set ensures the threshold actually corresponds to the expected error rate for routing decisions.

Extractions with low model confidence AND extractions from ambiguous or contradictory source documents. Both signal cases where automated acceptance carries significant risk.

Source URL · document name · relevant excerpt · publication/collection date. These travel with each claim through synthesis — they cannot be reconstructed after the fact once stripped.

Annotate both values with their sources and dates — do not select one arbitrarily. The coordinator decides reconciliation. The synthesis agent surfaces the conflict; it does not resolve it.

A Q1 figure vs. a Q4 figure may be a temporal update, not a factual contradiction. Without dates, the coordinator cannot distinguish a time-series difference from a genuine conflict.

Financial data → tables. News/qualitative findings → prose. Technical findings → structured lists. Converting everything to one format loses structure or nuance from the original source type.

A summarization step strips source URLs, document names, and excerpts from findings. The final report presents claims with no traceability — downstream consumers cannot verify or cite sources.