Confidence: medium. Evidence: controlled study and case study. Last substantive change: 2026-07.
This subsystem owns what the agent knows and how it comes to know it: the context it loads, the memory it keeps, and the reusable skills it draws on.
The conclusion
Keep durable truth in versioned artifacts and retrieve it progressively, and treat context files and skills as software supply-chain components with owners, tests, freshness, and provenance. Context design reliably changes outcomes. Skill ecosystems are promising, but they introduce drift and a real attack surface.
How the thinking got here
Giant prompts gave way to repository maps and just-in-time context, then to persistent sessions with externalized state, then to trainable and curated skill libraries. Along the way it became clear that a skill is not a harmless snippet: it is a dependency, and dependencies rot, drift, and can be poisoned.
Credible alternatives, and when each is right
| Approach | Right when |
|---|---|
| Full-context loading | small codebases |
| Retrieval | large corpora, targeted needs |
| Summaries | long histories that must be compressed |
| Event-sourced memory | auditability and replay matter |
| Clean-slate subagents | a task benefits from a fresh, distilled brief |
| Skills and code APIs | recurring, well-scoped capabilities |
Where it fails and what we still don't know
Failures include memory poisoning, stale context that silently misleads, and skill drift that breaks a contract the factory relied on. Evidence is strong that context design matters and contested on which document structures actually improve adherence; a notable result is that within-session decay, not file structure, is the measurable effect. Open questions include truth reconciliation, stale-context detection, skill dependency graphs, and safe forgetting.
What would change our mind
A standard for skill provenance and freshness that measurably reduces drift and poisoning would make the skill supply chain trustworthy by default.