Domain 5 is where agents meet the messy realities of production: conversations that run long enough to lose their own thread, failures that ripple across multiple agents, codebases too large to hold in one context window, and outputs that must stay trustworthy when they blend many sources. The common thread is deliberate management of information — what to keep, what to compress, how to move it between agents, and when to hand it to a person.
The first half of the domain is about context and reliability. Long interactions degrade in predictable ways: numeric values and stated expectations get flattened into vague summaries, and findings buried in the middle of a large input are more likely to be dropped than those at the edges. The countermeasures are structural — pulling transactional facts into a persistent “case facts” block that rides in every prompt, trimming tool results to the few fields that matter, placing key findings at the top of aggregated inputs, and delegating verbose exploration to subagents while the main agent keeps a clean, high-level view. Reliability adds the matching discipline for failure: retry transient errors, fail fast on bad data, and escalate business-logic cases with a structured summary rather than a raw transcript. Error reporting must be honest and specific, distinguishing a timeout that warrants a retry from a valid empty result, because generic statuses like “search unavailable” strip the coordinator of the context it needs to recover.
The second half turns to human oversight and provenance. Aggregate accuracy metrics can hide poor performance on a specific document type or field, so the domain emphasizes stratified sampling, field-level confidence scores calibrated against labeled validation sets, and validating accuracy per segment before automating high-confidence extractions. Confidence is treated as a routing signal, not an absolute truth — the model tends to over-report it — so ambiguous or contradictory sources are routed to limited human review where they matter most. Finally, multi-source synthesis is where attribution quietly breaks: summarization steps compress findings and lose the mapping from claim to source. The fix is to require structured claim-source mappings that downstream agents preserve and merge, to annotate conflicting statistics with their sources instead of silently picking one, and to carry publication dates so temporal differences are not mistaken for contradictions.
These skills reinforce each other. Good context management keeps a long exploration coherent; structured error propagation lets a coordinator recover instead of collapsing on a single subagent failure; calibrated confidence routing directs scarce human attention to the cases that need it; and provenance discipline keeps a synthesized report honest about what is well-supported and what is contested. Use the objectives above as a checklist: if you can explain why quality degrades gradually rather than failing outright, why escalation carries a summary and not a transcript, and why claim-source mappings must survive every summarization step, you have the spine of Domain 5.
Official exam objectives
- Manage conversation context so critical facts survive long interactions, countering progressive-summarization loss and the lost-in-the-middle effect.
- Extract transactional and structured issue data into persistent context layers, and trim verbose tool outputs before they accumulate.
- Design escalation and ambiguity-resolution patterns with explicit criteria, honoring direct requests for a human and asking for clarification on multiple matches.
- Implement error-propagation strategies that return structured error context and distinguish access failures from valid empty results across multi-agent systems.
- Explore large codebases with subagent delegation, scratchpad files, and structured state exports for crash recovery, avoiding context degradation.
- Design human-review workflows with stratified sampling and field-level confidence calibration, validating accuracy by document type before automating.
- Preserve information provenance with structured claim-source mappings, annotating conflicts and temporal data during multi-source synthesis.