Domain 5 · 15% of the exam
Context Management & Reliability
Manage conversation context, design escalation and ambiguity-resolution patterns, propagate errors across multi-agent systems, explore large codebases without context degradation, calibrate human review, and preserve provenance in multi-source synthesis.
5.1 Context Window Management
Hard facts (amounts, dates, order numbers, statuses) belong in a persistent structured case-facts block included in every prompt, outside the prose history that progressive summarization degrades.
5.2 Escalation & Ambiguity Resolution
Escalate on explicit human requests, policy gaps, and lack of meaningful progress — never on sentiment or a self-rated confidence number — and ask for more identifiers instead of guessing among multiple matches.
5.3 Error Propagation in Multi-Agent Systems
When a subagent fails, return structured error context — failure type, attempted query, partial results, alternatives — to the coordinator so it can recover intelligently, instead of a generic status, a silent success, or killing the whole workflow.
5.4 Codebase Exploration & Context Degradation
Long exploration sessions degrade — the model starts citing "typical patterns" instead of the specific classes it found — so offload verbose discovery to subagents, persist key findings in scratchpad files, and summarize before spawning the next phase.
5.5 Human Review & Confidence Calibration
Aggregate accuracy hides poor performance on specific document types and fields, so segment accuracy by type and field before automating, calibrate field-level confidence on labeled data, and use stratified random sampling to keep measuring high-confidence outputs.
5.6 Information Provenance & Multi-Source Synthesis
Carry claim-to-source mappings through every synthesis hop so citations survive, and when credible sources conflict, annotate both values with attribution and dates rather than arbitrarily collapsing them into one.