Analects: Agent-as-a-Model — An AI Coding Agent on Two DGX Sparks
I built Analects as an AI coding agent that runs as a standard OpenAI-compatible API. Point Continue.dev, Cursor, or curl at it — it handles the full agent loop server-side: planning, file editing, code search, web search, long-term memory, and code testing. You just get back a chat completion and fully built code or files in your workspace.
Two-Spark setup: Spark 1 runs the agent + 11 supporting services (router, vector DB, knowledge graph, web search, workspace sync from GitLab/GitHub). Spark 2 is dedicated to coder inference — Qwen3.5-35B on vLLM, ~57 tok/s with speculative decoding.
Built on Meta/Harvard’s Confucius research, extended with expert routing, a two-LLM architecture (big coder + small note-taker that learns from past sessions), and workspace indexing.
Self-contained: ./setup.sh setup + docker compose up -d starts everything.
Still early — under active development, things break, requires two Sparks (or equivalent ~128GB+ VRAM). But it works and it’s fun to hack on.
GitHub: GitHub - seli-equinix/analects: Analects: Agent-as-a-Model for code — Built on CCA/SWE-bench research from Meta · GitHub Docker Hub: hellohal2064/analects - Docker Image
PRs and feedback welcome, or help building it :)