context-window-stewardship

Context rot degrades AI agent performance during long-running sessions when token budgets overflow from undisciplined information loading, causing subagents to lose reasoning quality and cross-contaminate isolated contexts.

Manage context window token budgets by applying a 3-stage loading hierarchy that isolates essential triggers in Stage 1, loads the full body in Stage 2, and fetches on-demand resources in Stage 3 to prevent cross-talk.

You need context window stewardship when running long AI sessions, multi-agent orchestration, or memory-constrained deployments where unmanaged context loading causes token budget overflow and reasoning degradation.

Implement the 3-stage loading hierarchy by keeping Stage 1 triggers in SKILL.md, loading the Stage 2 full body on task match, and fetching Stage 3 references, scripts, and assets on demand for specialized workflows.

Yes, context window stewardship supports long-running agent sessions with multiple subagents by enforcing isolated contexts through the 3-stage loading hierarchy, preventing cross-talk and preserving token budgets.

Context window stewardship requires disciplined adherence to the 3-stage loading hierarchy; skipping stages or loading Stage 3 resources prematurely will still cause token budget overflow in memory-constrained deployments.