Show HN: Autonomy – Self-Harness/Self-Directed AI Agent Core Under Development
3 days ago
- #tool-integration
- #ai-workflow
- #autonomous-agents
- AgentLoop is the core executor of Autonomy, designed as an 'autonomous task executor' that operates without human intervention to achieve goals, autonomously selects tools, and acts, learning from results.
- Each Turn involves selecting skills via ProcedureSkillLibrary, generating action paths via LLM proposals and RecipeEngine, ranking them with a beam width of 3 using CandidateSelector based on scoring criteria, and executing through ActionGateway with approval policies.
- Autonomy supports interactive and batch modes, with all events logged in AutonomyStore using run_id for traceability and replay via Event Sourcing.
- Stopping conditions include ACHIEVED (goal met), BLOCKED (unable to proceed), NO_CANDIDATES, APPROVAL_DENIED, MAX_STEPS_REACHED, or FAILED to prevent infinite loops.
- LearningLoop reviews runs post-completion; if achieved with at least 2 successful steps, it automatically creates ProcedureSkillDraft and LearningProposal for approval.
- RecipeEngine generates SHA-256 fingerprints for successful actions; when an action succeeds ≥2 times, it becomes a Recipe for direct reuse without LLM re-reasoning.
- CuratorDaemon asynchronously merges or subsets skills post-run to prevent skill bloat, using confidence thresholds (0.95 for duplicates, 0.90 for subsets).
- Toolset includes TOOLSET_CATALOG with core tools (file, terminal, search, skills) and optional project/browser tools, plus memory, delegate, cronjob, and computer_use.
- Security features include ActionGateway approval policies with RiskLevels (LOW auto-approved, MEDIUM requires approval), redaction of sensitive data in logs, and deterministic outcome evaluation.
- Setup involves cloning the repository, installing dependencies, configuring models, and using autonomy tui for a terminal interface that streams natural language to AgentLoop.
- Supported providers include ollama (local) and 8 OpenAI-compatible services (e.g., NVIDIA, DeepSeek, Kimi), with API key management and connectivity checks via autonomy doctor.
- MCP tools are integrated via mcp_servers.yaml and managed by ToolRegistry, while memory tools provide persistent workspace memory and database tools support SQL with natural language translation.