context-engineering

To improve agent output quality during codebase onboarding, load a hierarchical context starting with project rules files, then specs, relevant source files, and error output to reduce hallucinations and drift.

The best way to manage agent context is using a search evidence workflow that forces target verification before edits, applying selective, task-scoped context loading to include only necessary files.

AI agents silently guess due to missing or overly broad context; you can surface these spec-vs-code conflicts and missing requirements by applying explicit uncertainty handling patterns.

Set up selective context loading by prioritizing rules files first, then architecture specs, followed by specific source files and failing test output, explicitly excluding irrelevant data to maintain conventions.

You need to engineer agent context when starting sessions, switching workstreams, onboarding to unfamiliar codebases, or troubleshooting declining results and degraded AI performance during iteration.

Handle missing requirements and outdated context by curating exactly what the agent sees through hierarchical context selection and explicit conflict handling, which reduces hallucinations and maintains task focus.