context-debugging

Agent failures often stem from incorrect, missing, or conflicting context rather than code bugs. Diagnosing the assembled context—system prompt, tools, and retrieved docs—against intended instructions isolates whether buried directives or context overflow cause the hallucination.

Debugging agent regression requires verifying what the model actually receives post-change. Comparing actual context assembly against intended definitions identifies missing directives or tool definition problems introduced by recent modifications.

Fixing conflicting instructions requires classifying the agent failure into context failure modes, then applying category-specific remediation. Routing buried or ambiguous directives through targeted context verification resolves instruction conflicts systematically.

Diagnosing excessive context involves a context-first triage to verify system prompt, tools, and history volume. Classifying the issue as context overflow matches it to focused remediation steps that reduce assembly bloat.

Yes, diagnosing unexpected output switches applies when agents regress from expected behaviors. Verifying actual versus intended context and checking for instruction conflicts resolves unexpected alternative outputs without modifying harness code.

Context debugging does not resolve genuine reasoning failures or infrastructure and harness code bugs. It applies exclusively to agent failures caused by incorrect, missing, ambiguous, conflicting, or excessive context assembly.