context-engineering

Agent failures frequently originate from incorrect information in the model context, specifically missing, stale, wrong, or overwhelming data. Diagnosing these context problems requires analyzing the prompt stack and tool-result lifecycle to pinpoint the root cause.

Debug missing or stale context by walking the Missing vs Stale vs Wrong vs Overwhelming decision tree to isolate the failure mode. Then identify the exact context layer—system prompt, persistent memory, or injected skills—and apply the targeted fix to restore context freshness.

Measure LLM context quality by tracking injection precision, injection recall, context utilization, and a freshness score. These metrics quantify the health of your context payload and help prevent context pollution across agent sessions.

Use subagent delegation to manage agent context when you need cleaner main-agent context during heavy investigations. Delegating context-heavy research to a subagent prevents overwhelming information from polluting the primary prompt stack and maintains reasoning accuracy.

Prevent context pollution when injecting skills by applying frequent intentional compaction and validating the context pipeline against a debugging checklist. This ensures injection precision and stops stale or wrong information from overwhelming the agent prompt stack.

Yes, context engineering works for troubleshooting persistent memory by applying a five-layer context stack model. It specifies how always-loaded rules and persistent memory fail over time, providing validation checklists to ensure memory freshness and reduce drift.