Atlas · agents
Harness
The program around the model that builds context, runs tools, enforces rules and keeps state between sessions.
See it working: Plate I, Plate II
In plain words
Claude Code, Codex and Cursor are harnesses. The model produces text. The harness turns that text into actions on your computer and decides what the model gets to see.
How it works
A harness assembles the system prompt, tool descriptions, instruction files and conversation into one context, calls the model, and acts on its reply. It also enforces permissions and runs hooks, which the model cannot override.
Going deeper
Long-running work needs state that outlives one context window. Anthropic describes harnesses that leave progress files for the next session to pick up. Agent Atlas does the same with its roadmap, journal and handoff files.
Sources
- How Claude Code works Anthropic
**gather context**, **take action**, and **verify results**
- Effective harnesses for long-running agents Anthropic
while leaving clear artifacts for the next session.
The very first agent session uses a specialized prompt that asks the model to set up the initial environment: an init.sh script, a claude-progress.txt file that keeps a log of what agents have done
After some experimentation, we landed on using JSON for this, as the model is less likely to inappropriately change or overwrite JSON files compared to Markdown files.
Quotes were checked against the live pages on 2026-11-08.