Self-Improving Agents Through Closed-Loop Learning
GitHub pip install cannyforgeSkills are defined declaratively via SKILL.md files. The engine handles execution, error detection, pattern learning, rule application, and lifecycle automatically.
Try It NowDefine skills via SKILL.md files following the AgentSkills.io spec. No Python subclassing required.
Monitors errors, detects patterns, generates rules. Triggers automatically when 2+ uncovered error types appear.
Rules that underperform are demoted to PROBATION, then DORMANT — never deleted. Resurrection when errors resurface.
Works with LangChain and CrewAI. Wrap any skill as a tool for your existing agent.
Three-tier execution: custom handler → LLM multi-step loop → template fallback. Your choice.
JSON file or SQLite backend. Your data stays on your machine.
# Install
pip install cannyforge
# Run the demo
cannyforge demo
# Or use in code
from cannyforge import CannyForge
forge = CannyForge()
result = forge.execute("Write an email about the 3 PM meeting")
# First run: False (TimezoneError logged)
# After learning: True (rule prevents the error)
print(result.success, result.output)