AtlasAgent
An auditable AI agent control plane with evidence-backed memory, a governed tool runtime, checkpoint-based recovery and a companion engineering tutorial, aimed at answering where a fact came from and what an agent did unattended.
AtlasAgent is a control plane for building a vertical AI agent that plans a task, executes it step by step, calls tools, observes results, and answers with evidence it can point back to, rather than only a chat shell wired to a language model.
Its memory system requires facts to pass a write gate that records source, scope and expiry before they reach the model, then applies decay, consolidation and conflict resolution over a typed memory graph. Its retrieval layer rewrites queries, fuses results from multiple queries, retrieves parent documents, and reranks with a language model with a lexical fallback, with every citation carrying a confidence score and every query leaving a retrieval trace. Its tool runtime gates every call through a risk tier, an approval step, an idempotency check, a timeout, redaction and an artifact record before the handler runs, with result caching, dependency-aware batching, retries with fallback, and per-step budgets. Before a task is marked done, an acceptance command must exit successfully, a scope audit checks that only the files the plan declared were touched, and a language model reviews test coverage.
The project ships with FastAPI, a Next.js progressive web app and a terminal interface, plus a 55-chapter tutorial that rebuilds the whole system from an empty directory.