Agent Learning Loop

Track task outcomes, accumulate experience, and improve execution over time. Turns every task into a learning opportunity.

Commands

Run via: python3 scripts/agent-learning-loop.py [options]

Core Commands

• record --task --approach --outcome [--duration ] [--notes ] — Record a task execution and its outcome
• suggest --task — Get the best known approach for a task type (based on history)
• patterns — Analyze and display discovered task patterns
• stats — Show overall learning metrics (success rate, improvement trend, top patterns)
• review --days — Review recent task history for learning opportunities
• prune --older-than — Archive old entries, keeping distilled learnings

Options

• --category — Task category (coding, research, writing, ops, finance, etc.)
• --json — JSON output
• --min-confidence <0-100> — Minimum confidence for suggestions (default: 60)

How It Works

1. Record Phase

After each significant task, record what was done and how it went:

agent-learning-loop.py record \
  --task "Deploy Cloudflare Pages site" \
  --approach "wrangler pages deploy with --commit-dirty flag" \
  --outcome success \
  --duration 45 \
  --category ops

2. Pattern Recognition

The system clusters similar tasks and identifies:

• Winning approaches — What works consistently for this task type
• Failure modes — Common reasons tasks fail
• Time patterns — How long similar tasks typically take
• Category trends — Which categories are improving

3. Suggestion Phase

Before starting a familiar task, ask for the best approach:

agent-learning-loop.py suggest --task "Deploy website to Cloudflare"

Returns the highest-scoring approach with confidence level.

Data Storage

• memory/learning-loop.json — Task history and pattern database
• memory/learning-patterns.json — Extracted patterns and winning approaches

Pattern Matching

Tasks are matched using:

1. Keyword overlap — Shared significant words between task descriptions

2. Category match — Same category boosts match score

3. Outcome weighting — Recent successes weighted higher than old ones

4. Decay — Older entries contribute less (30-day half-life)

Cron Integration

Schedule a daily review to extract patterns:

agent-learning-loop.py review --days 1

References

• See references/pattern-extraction.md for the pattern matching algorithm details