Progressive Automation Practice¶
A graduated approach to autonomous agent loops that builds confidence and prevents costly failures.
The problem¶
Jumping from zero automation to overnight unattended runs is high-risk. A single misconfigured loop can burn tokens, produce wrong output, or silently diverge from intent. Progressive practice builds verified confidence at each stage.
The four stages¶
Stage 1: Single iteration (manual)¶
Run one pass of any automated workflow manually. Verify output. Understand what the agent does at each step.
Graduate when: You've run 3+ single iterations with correct output.
Stage 2: Short attended loop (2-3 iterations)¶
Let the loop run 2-3 iterations while you watch. Interrupt if anything looks wrong.
Graduate when: 3+ short loops complete without needing intervention.
Stage 3: Longer attended loop (5-10 iterations)¶
Extend the loop duration. Check in periodically but don't watch every step.
Graduate when: The loop handles edge cases and errors gracefully without your input.
Stage 4: Unattended loop¶
Only after stages 1-3 have built verified confidence. Set up proper guardrails:
- Max iterations: Always set a hard limit (ralph-wiggum-loop defaults to 5)
- Token budget: Monitor token usage — runaway loops are expensive
- Output review: Always review output after unattended runs
- Rollback plan: Know how to undo what the loop did (git reset, file backups)
Common mistakes¶
| Mistake | Prevention |
|---|---|
| Skipping to Stage 4 | Always start at Stage 1 for new workflows |
| No iteration limit | ralph-wiggum-loop enforces max 5; never remove this |
| Ignoring token costs | Check metrics/verification-loops.json after each run |
| Trusting completion claims | The Iron Law: NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE |
| Running overnight without guardrails | Set token budget alerts, use max iterations, review output next morning |
Integration with toolkit skills¶
/vt-c-ralph-wiggum-loop: Convergence-based verification with built-in safety limits/loop: Native interval-based scheduling — pair with ralph-wiggum-loop for verified convergence- Metrics: Check
metrics/verification-loops.jsonto track false claim prevention and loop efficiency