Goal Mode — Writer/Checker Verification Loop for AI Agents

A pattern for running AI agent tasks with a separate writer and verifier. The writer produces output. The checker verifies it against explicit conditions. If the checker fails, the writer iterates. If blocked, it reports honestly.

Core insight: The model that produces the output should not be the same instance that decides it's done.

How it Works

Goal → Writer → Output → Checker → PASS? → Report
                              ↓
                           FAIL (< 3 retries)
                              ↓
                         Writer fixes → Check again
                              ↓
                           FAIL (3 retries) → Report gap
                              ↓
                           BLOCKED → Report gap

Activation Gates

The loop only activates when ALL of these pass:

1. Persistent side effect — output writes to code, Notion, API, file, or DB
2. Objectively verifiable — "done" can be checked with a test, tool call, or file read
3. Error cost > token cost — undetected failure would be expensive
4. Worth the overhead — task would take 5+ minutes without a loop

Never enters the loop for: answers, summaries, drafts, exploratory work, informational cron jobs, or tasks under 2 turns.

Included

• goal-mode.skill.md — The operational skill. Load this in an agent to run the writer/checker loop.
• goal-drafting.skill.md — Reference template for writing verifiable finish line conditions.

Known Failure Modes (Mitigated)

Failure	Mitigation
Checker false positive	Different model families for writer vs checker
Checker false negative	Structured JSON output + specific feedback
Verbosity bias	Include length constraints in conditions
Loop instability	Hard cap of 3 retries, then escalate
Blocked items	Must report "Cannot verify" — not allowed to guess
Checker limited to text	Checker has tool access (files, terminal, search)

Production References

This pattern is used in production at:

• Spotify — hourly playlist moderation (writer + evaluator)
• Uber — 15-min notification pipeline (writer + verifier, JSON schema reduced false positives from 5% to 0.3%)
• Datadog — daily monitor comments (switching model families reduced errors by 60%)
• Stripe — per-PR code review (checker with confidence thresholds)
• Airbnb — 5-min listing moderation (evaluator with length normalization)
• Anthropic docs — tool-use evaluation loop as canonical pattern
• LangGraph — "reflect" node pattern for separate evaluation
• CrewAI — native verifier agent with retries

License

MIT