V
All projects
Case study

Atreus

6+ coordinated agents · semantic memory · self-hosted

Diagram of the Atreus orchestration system: an orchestrator directing six specialist agents for planning, building, validation, scheduling, and archiving, with persistent memory and verification gates

What it does

Atreus runs on my own server on top of OpenClaw, and it exists so the execution layer around my projects stops needing me in the room. I describe an outcome over a chat interface and it creates the branch, writes or edits the files, runs the build, checks the result, pushes, and comes back with a preview link for me to approve or send back with notes.

How it is structured

Six specialist agents sit behind one orchestrator, split by job rather than by model: planning, building, validation, scheduling, and archiving. Persistent memory is maintained on a scheduled overnight pass, so the system stays oriented across sessions instead of needing context re-established every time.

The verification gate

The design decision worth keeping is the verification gate. Early on it would report a task complete when it had failed, because a plausible success message is what success usually looks like and nothing was checking. Now nothing counts as done without evidence from the artifact itself: the actual build output, a live check of the deployed page, a grep for the content that should be there. An agent's self-report is not evidence, which is the one lesson from this that transfers to any system where an agent grades its own work.

Key results

  • 6+ coordinated agents
  • Semantic memory search
  • Autonomous web dev pipeline

Stack

PythonDockerAI OrchestrationVector Database