Atlas (formerly Arch) is an autonomous agentic framework built on LangGraph that learns, refines, and executes specialized skills using Gemini 2.5.
- Code Walkthrough: A comprehensive guide to the Atlas architecture, nodes, and first-class skill tools.
- Latest Progress: Real-time update on the latest features and stability fixes.
- Install:
pip install -e . - Configure:
Set
GOOGLE_API_KEYandSUPABASE_URIin your.env. - Run:
python -m skill_agent.cli
- First-Class Skill Tools: Dynamically loads @tool decorated scripts from the
skills/directory. - Gemini 2.5 Tiering: Uses
flashfor planning/execution andprofor reflection. - Surgical Refinement: Preserves work context when looping back for refinements.
- Persistent Checkpointing: Integrated with Supabase/Postgres for long-running task persistence.
- Strict Sequencing: Enforces turn-taking rules to ensure clean LLM interactions.