ATLAS In development · Early access list open
The Atlas systemOverview

What Atlas is

Not another agent.
The one that coordinates them.

Atlas is designed so that you give it a goal and it plans the work, hands parts to the right AI models, has the result checked, and brings it back to you to accept, with a record of everything that happened.

01 / How it holds togetherThree jobs

One mission.
Three jobs.

Coherence

Keep the goal straight

Design goal: the goal, decisions and what "done" means should survive each handoff between models.

Policy

Ask before acting

Design goal: anything consequential is checked against what Atlas is allowed to do before it happens.

Each agent has a defined role and a list of commands it's allowed to run.

How Atlas governs AI work ↗

Record

Keep the receipts

Design goal: who did what, with which model, and what checked it, kept in one history you can inspect.

A record of what shipped is kept by the team today; Atlas keeping that record itself is planned.

02 / RolesPlain names

Who does
what.

Planner

Plans

Turns your request into a plan you approve.

Builders

Build

Several models work in parallel.

Reviewer

Checks

A separate reviewer checks the change.

Hermes

Supplies models

In testing: we're testing Hermes to bring in local and open models on demand.

Example mission

How a mission flows.

Everything about one mission stays in one place. The team thread shows each step as it happens, and, in the design, you can message the planner, a builder or the reviewer directly.

03 / IntegrationsWhat it connects to

What Atlas
runs on today.

Models

Claude Code, Codex, Antigravity, Grok

On Atlas's own work, builders and reviewers come from different providers, assigned per task. Model · effort: the Codex lane is routed by model and effort for each task, and each review records the model and effort requested and observed. Showing model and effort on every assignment is planned.

Code

GitHub

Atlas's own work lands through GitHub pull requests, tested by GitHub Actions CI.

Planning

Issue tracker

Atlas tracks its own work in Linear. Starting missions from your issues is planned.

Claude and Claude Code (Anthropic), Codex (OpenAI), Gemini and Antigravity (Google) and Grok (xAI) are trademarks of their respective owners. The symbols shown are original Atlas designs, not provider logos; Atlas is not affiliated with or endorsed by these companies.

04 / StatusHonest status

Where it
stands.

Atlas is in development.

The coordination core is being built and tested on Atlas's own code. Running a full mission without hand-holding is the next milestone. What's on this page describes how the system is designed to work. The mission thread above is an example.