Summary

Long-running agent systems get more reliable when context selection is explicit, inspectable, and policy-constrained.

Key points

• Per-turn context manifests make selection and exclusion visible.
• Governed retrieval reduces token waste and improves debuggability.
• Default-deny or constrained retrieval policies matter when systems become more agentic.
• This pattern is a safer evolution path than broad "everything is a file/action" ambition.

Operational relevance

This is one of the strongest remaining extractable patterns from QUA-260 because it links context engineering, governance, and runtime reliability.

Open questions

• What should a minimal context manifest include for current workflows?
• Which retrieval surfaces should remain gated until better policy coverage exists?

Sources

• wiki/sources/everything-is-context-critical-assessment.md
• wiki/sources/openai-harness-engineering.md