← All products
🔄

Self-Improving Agent Kit

Your AI agent repeats the same mistakes. This kit gives it a 6-component feedback pipeline that turns corrections into tested rules, tracks whether those rules hold, and prevents self-improvement from becoming runaway behavior.

✓Now with promotion gates and rollback rules.
Updated May 11, 2026 · Added bounded feedback loops, rule rollback, and promotion gates so learning does not become runaway behavior.

Paid Digital Thoughts subscriber?

Yearly subscribers: all products free ($1150.94+ value)

Monthly subscribers: 1 free product per month — subscribe from $29.99/mo

Claim your free copy →
$49One-time purchase.

What changed in this update

  • ✓Added bounded-feedback-loops.md
  • ✓Added rule-rollback-playbook.md
  • ✓Added CHANGELOG.md and quickstart safety notes

Best for

  • +Agents that receive corrections and need to stop repeating them
  • +Builders who want measurable behavior improvement without fine-tuning

Not for

  • -Autonomous rule rewriting without human review for sensitive workflows

What you get

+Signal collection — append-only JSONL log of tasks, corrections, approvals, and errors
+Correction detection — pattern-matching categorizer (verbosity, style, wrong approach, etc.)
+Style learning — banned phrases, voice patterns, verbosity ratios
+Performance tracking — 0-100 score from task success rate, correction rate, efficiency
+Error registry — structured database for systemic errors and resolutions
+Lessons system — behavioral rules with graduation (max 50 active)
+Bounded feedback loops — limits for new rules, signal thresholds, and test windows
+Rule rollback playbook — retire rules that make the agent worse
+All pure Python, no external dependencies, no API keys required

Package includes

  • •5 standalone Python scripts (signal_collector, correction_detector, style_learner, performance_tracker, error_registry)
  • •setup.sh — one-command installation
  • •CLAUDE.md integration snippet
  • •Templates (lessons.md, error-registry.json, signals.jsonl)
  • •bounded-feedback-loops.md and rule-rollback-playbook.md
  • •Real examples (correction report, sample signals)
  • •Comprehensive guide (10 sections, architecture to production)
  • •QUICKSTART — working pipeline in 10 minutes
  • •CHANGELOG.md

FAQ

Does this need an LLM to analyze corrections?

No. Correction detection uses Python regex pattern matching — fast, free, deterministic, and ~90% accurate for common categories.

How is this different from just updating CLAUDE.md manually?

Manual updates are guesswork. This kit uses data: it counts corrections, finds patterns, proposes rules, and tracks whether your changes actually reduce correction rates.

How long until I see improvement?

Week 1: signals accumulate. Week 2: patterns emerge. Week 3: rules applied, score starts climbing. Month 2: agent rarely repeats old mistakes.

Related reading on Digital Thoughts:

Secure checkout by Stripe. Instant download + guided Claude Code setup.