On this page
/ docs · architecture

Architecture

MAARS is easiest to understand as a deterministic Python runtime coordinating a few open-ended agent stages through a file-based session store.

01What this project is

02System boundaries

The architecture becomes much easier to reason about once the boundaries are explicit.

03Stage contracts

Each stage reads a small set of persisted inputs and writes a small set of persisted outputs.

04End-to-end data flow

The main pipeline is a chain of files on disk, not a chain of in-memory messages.

05Runtime vs agent responsibilities

Python decides control flow. Agents do the open-ended work inside each step.

06Session storage and recovery

One run becomes one directory under results/, so sessions are inspectable and recoverable.

07Code map

The codebase mirrors the architecture: entry, orchestration, stage logic, tools, then storage.