ATLAS¶
A failure-mode taxonomy layer for agents: reflect at meaningful boundaries, catch recurring mistakes, and learn from traces.

ATLAS gives an agent a structured way to notice recurring mistakes at runtime, record evidence, and generate or refine task-specific failure-mode taxonomies from completed traces.
Runs start from MAST — the Multi-Agent System failure Taxonomy from "Why Do Multi-Agent LLM Systems Fail?" (Cemri et al., 2025), shipped as a built-in 14-code adaptation. At configured gates the agent reflects on its recent trajectory (Observe → Correlate → Map → Decide), a blocking final gate runs before submission, and completed traces feed taxonomy generation and refinement. ATLAS is not a task solver; your harness keeps owning model execution.
Quickstart¶
Install:
python -m pip install "git+https://github.com/multi-agent-systems-failure-taxonomy/ATLAS.git"
Create atlas.json in the project that will run the agent:
{
"version": 1,
"trace_output": "./atlas-program",
"atlas_model": "gpt-5"
}
Then choose the integration that matches your pipeline:
| Use case | Command | Full docs |
|---|---|---|
| Claude Code project | atlas-claude-install --project-dir . --config atlas.json |
Claude Code |
| Codex project | atlas-codex-install --project-dir . --config atlas.json |
Codex |
| One LLM call from a script | atlas-single-run --config atlas.json --task "..." --model gpt-5 |
Single LLM |
| Existing trace folder | atlas-import-traces --config atlas.json --traces ./traces |
Taxonomies |
| Your own harness | from atlas_runtime import start_session, ... |
Pipeline integration |
Check the setup:
atlas-doctor --config atlas.json
Where to start¶
- New to the terminology? Concepts defines the vocabulary and the runtime loop.
- Want to see real output first? An example run shows the reflections, the final gate, and the dashboard.
- Ready to wire it up? Getting started is the 5-minute
path; Configuration reference covers every
atlas.jsonfield.