executive summary: amp thread analysis
corpus: 4,656 threads | 208,799 messages | 20 users | may 2025 – jan 2026
top 5 findings
| # | finding | impact |
|---|---|---|
| 1 | file references in opener (@path) | +25pp success (66.7% vs 41.8%) — strongest single predictor |
| 2 | approval:steering ratio > 2:1 | 4x success vs <1:1 — ratio predicts thread health |
| 3 | 26-50 turns is optimal | 75% success vs 14% for <10 turns — most threads die too early |
| 4 | steering = engagement, not failure | 60% resolution in steered threads vs 37% unsteered |
| 5 | confirm before action | 64% of steerings correct premature action (“no”, “wait”) |
top 5 recommendations
| # | recommendation | implementation | expected impact |
|---|---|---|---|
| 1 | include file references in opening message | zero effort — type @path/to/file | +25% success rate |
| 2 | approve explicitly after successful steps | type “good”, “ship it”, “yes” | maintains 2:1 ratio, 4x success |
| 3 | stay past 10 turns on meaningful tasks | don’t abandon prematurely | +61pp for 26-50 vs <10 turns |
| 4 | add confirmation gates to AGENTS.md | agent confirms before tests/commits/scope changes | -64% steering interventions |
| 5 | use oracle at planning, not rescue | invoke early for architecture, not late when stuck | prevents frustration spiral (46% of frustrated threads used oracle as last resort) |
expected impact
conservative estimate: implementing all 5 recommendations could move team resolution rate from current 44% to 60%+ based on observed correlations.
individual user improvements:
- verbose_explorer: +26pp possible by staying 30+ turns and avoiding evening sessions
- feature_lead: -20pp handoff rate by spawning subtasks vs abandoning
- team average: 2:1 approval discipline alone correlates with 4x success
key insight
steering is a feature, not a bug. the counterintuitive finding: threads WITH user steering resolve at 60% vs 37% for threads without steering. steering indicates engagement, not failure. the problem is not steering itself, but:
- consecutive steerings (doom spiral forming)
- steering without subsequent approval (no checkpoint established)
- ratio inversion (<1:1 approval:steering = danger zone)
implementation roadmap
| phase | action | owner |
|---|---|---|
| immediate | update AGENTS.md with confirmation gates | team |
| week 1 | share quick-wins.md with all users | lead |
| week 2 | implement thread health monitoring (ratio tracking) | tooling |
| ongoing | review approval:steering ratios in retros | team |
synthesized from 87 insight files | 2026-01-09