| name | run-comparison |
| description | Run prompt-only vs agent-loop vs background-agent comparison on the same task and log structured metrics. Use for workshop demos, A/B evidence, or cohort measurement. |
Run Comparison (A / B / C)
Hold the task constant. Vary only the method. Log metrics for side-by-side evidence.
Methods
| Arm | Label | Human role | Verifier |
|---|
| A | Prompt-only | Closes the loop — judges output manually | None (human eyeballs) |
| B | Agent loop | Reviews at green; @loop-engineering + rules | Machine-checkable |
| C | Background / Cloud | Reviews the PR when agent finishes | Full verifier + budget |
Procedure
1. Pin the task
Use a versioned task id from the corpus (e.g., oms-status-race-001). Same seeded state, same verifier, same budget for B and C.
2. Run arm A (prompt-only)
- Open a fresh thread in Chat or Inline Edit — not Agent mode.
- Paste the task goal without mentioning the verifier command.
- Apply the suggested fix manually.
- Human runs the verifier and records pass/fail.
- Log: pass@1, human interventions, wall-clock, whether human ran verifier.
3. Run arm B (agent loop)
- Fresh Agent mode thread.
- Load loop rules:
@loop-discipline @loop-termination @loop-safety.
- Invoke
@loop-engineering with full task spec including verifier.
- Let the agent iterate until green, blocked, or budget.
- Log: pass/fail, iterations-to-green, human interventions, wall-clock, regression count.
4. Run arm C (background) — optional
:::caution VERIFY
Confirm current Background/Cloud Agent UX and CLI flags against cursor.com/docs before facilitating arm C.
:::
- Launch background agent with the same prompt + verifier spec as arm B.
- Human reviews the resulting PR.
- Run verifier on the branch; log same metrics as B.
5. Write metrics file
Append to loop-runs/<task_id>/<timestamp>.json:
{
"task_id": "oms-status-race-001",
"method": "B",
"pass": true,
"pass_at_1": false,
"iterations_to_green": 4,
"human_interventions": 1,
"wall_clock_seconds": 312,
"regression_count": 0,
"notes": "Agent fixed symptom first; regression caught by suite gate"
}
Facilitation notes
- Run A before B so learners feel the cognitive load difference.
- Report cost and wall-clock honestly alongside pass rate — credibility depends on it.
- Same facilitator, same model selection across arms when possible.
CLI harness
:::caution VERIFY
The one-command harness (cursor-agent -p ...) is V2 milestone. For V0 workshops, run arms manually and log to JSON. Re-verify CLI syntax at cursor.com/docs/cli.
:::
./scripts/loop-harness.sh --task oms-status-race-001 --method B