FullStack VC

Proposed · Context Vigilance WG

Context Vigilance

A practical framework for keeping AI agents performant and accurate — managing the context they read, what they remember, and what they're allowed to change.

Part of: Hack & Ship

All projects Edit on source

Why this exists

Most teams using agents are still in the "demo it once, marvel, then quietly stop" phase. The reason isn't model capability — it's context discipline. The agent that worked yesterday gets fed slightly different context today and produces noticeably worse output. Nobody has a name for what went wrong.

Context Vigilance gives that practice a name and a toolkit. We're not optimizing models. We're optimizing the context — what the agent sees, what it remembers, what it's allowed to do with what it learns.

What we're building

  • A vocabulary for the practice (context budgets, memory tiers, trust boundaries, review motions).

  • Working examples per role (an analyst's daily loop, an associate's diligence loop, a partner's review loop).

  • Diagnostic tools — when output drifts, what changed in the context?

Working group expectations

  • Bring an agent loop you actually run, not a hypothetical. We diagnose live.

  • Read each other's memory files. The vigilance is collective.