Skip to content

The Problem

Without managed startup, AI agent sessions suffer from six issues:

  1. Context waste — loading everything every session burns tokens on rules that aren't needed for the current task. A 3000-line rule set costs ~4K tokens on every session, even when you're just fixing a typo.

  2. Rule drift — facts, counts, and references go stale. Your project instructions say "58 rules" but the DB has 62. No one notices until a rule is missed.

  3. Startup chaos — CLAUDE.md says "read these files" but there's no enforcement. The agent skips files, partially loads context, or starts working before critical rules are loaded. Writing "you must read X" in a markdown file is documentation, not enforcement.

  4. Information scattering — learnings, decisions, and conventions accumulate in conversations and random files. When the session ends, they're lost. Without a routing mechanism, the knowledge base stays fragmented.

  5. Session amnesia — every new session starts from zero. The agent doesn't know what was done last time, what's in progress, or what's next. Users re-explain context every session.

  6. Unverified completion — the agent says "done" but the fix wasn't tested, the commit wasn't pushed, or the config wasn't validated. Generating output feels like completing the task, but the gap between "I did it" and "it's done" is where failures hide.

This architecture solves all six:

  • Tiered loading — load what's needed, defer the rest (problems 1, 2)
  • Structural gates — block tools until context is loaded (problem 3)
  • Drift detection — catch stale references automatically (problem 2)
  • Rule Zero — route scattered information to consolidated files at edit time (problem 4)
  • Session continuity — persistent backlog and session handoff across sessions (problem 5)
  • Self-verification — block exit until work is verified, not just narrated (problem 6)

See What It Does For You for the full capability map.