03 · ANDREA · AGENT OS

The agent that runs the workshop — under supervision.

Andrea is the operating layer across every division: persistent memory of company knowledge, durable graph execution for real work, and verified self-repair of real repositories. It never merges anything that doesn't pass the actual test suite.

RUNTIME STABLE · 380 TESTS GREEN

Capabilities

Persistent memory

SQLite-backed store with categories (episodic / semantic / procedural), confidence scoring, expiration and provenance. Andrea remembers what was done for whom — and why.

Durable graph runtime

Work expressed as graphs: parallel node scheduling, crash recovery that never re-runs completed nodes, and git-worktree isolation so experiments can't touch the real checkout.

Verified self-repair

The repair lifecycle is an ordinary graph — diagnose the failing tests, implement a fix in isolation, re-run the real command, and merge only through an explicit human authorization gate.

Permission engine

Every tool call classified AUTO / ASK / DENY with command-risk classification. High-risk actions require explicit approval via Telegram or terminal. The model cannot widen its own permissions.

Trust model

Autonomy is earned per action.

  • Read-only tool allowlists injected at runtime — the model cannot widen them
  • Mutation work happens in isolated git worktrees; the original checkout stays untouched
  • Authorization gates are explicit and recorded: approve / deny with evidence
  • Verification runs against real exit codes, never the executor's own conversation

andrea graph repair --repo ez-industries

baseline → 3 failing tests recorded

diagnosis → wrong operator in risk_analytics

implementation→ fix applied in worktree wt-7f31

tests → 282 passed ✓

verifier → rule + agent, real exit codes ✓

authorization → WAITING (telegram: /approve)

Providers

Grok (xAI) cloudOllama localDeepSeek thinking modeText tool-call normalization
NEXT: COMPUTER CONTROL · VOICE · VISION →