Autonomy

Autonomy is an open-source Python framework for building autonomous agents that plan, execute, and learn from their own runs without requiring a human to drive each step. The AgentLoop starts from a natural-language goal, pulls relevant ProcedureSkills from a library of 15 bundled skills, generates candidate action paths via a combination of LLM proposals and a RecipeEngine, scores them across five dimensions — evidence strength, purpose, risk, side effects, and penalty — then executes the top-ranked candidate through ActionGateway. Six explicit termination conditions (ACHIEVED, BLOCKED, NO_CANDIDATES, APPROVAL_DENIED, MAX_STEPS_REACHED, FAILED) ensure the agent cannot loop indefinitely.

The differentiating feature is the combination of ActionGateway and the post-run LearningLoop. Every tool call carries a RiskLevel and a verification plan before execution. After a successful run with at least two positive steps, LearningLoop automatically drafts a new ProcedureSkill at 0.85 confidence and generates a LearningProposal — which waits for review rather than self-applying. The RecipeEngine separately fingerprints successful actions via SHA-256 and promotes repeated patterns to reusable recipes, reducing LLM calls on familiar tasks over time.

The framework fits teams that need a verifiable, auditable agent for tasks like systematic debugging, TDD cycles, codebase documentation, or browser-based inspection — all covered by bundled skills. It fits less well when you need agents coordinating in parallel or a goal that branches into concurrent subgoals; the single AgentLoop design does not natively support multi-agent parallelism. Teams hitting that ceiling typically add an external orchestration layer, at which point they are managing two systems.

The toolset is extensible: file, terminal, search, and skills are enabled by default; browser (Playwright headless Chromium, 11 tools) and project tools (git, JSON, YAML) are opt-in; and external tools are importable via MCP through mcp_servers.yaml with MEDIUM risk classification applied automatically. Nine LLM providers are supported, including local Ollama and eight OpenAI-compatible endpoints, with an autonomy doctor command to verify endpoint connectivity.