Core Mental Model

Little Harness is a runtime boundary around an agent. The developer gives the agent instructions and capabilities; the harness supplies the working environment.

The important shift is that the developer does not define every directory, shell, artifact sink, and session store before the first useful agent works. Those are managed defaults. The developer only opts into app-specific persistence, exports, and policy.

developer config
-> tools + skills + optional Persistent Dirs
-> auto-created session
-> isolated workspace + artifacts + trace
-> assistant result + host-visible status

The Main Surfaces

Session is the private conversation workspace. It persists across turns in one conversation and is isolated from other sessions.

Session Files are files inside that workspace. They can be staged by chat messages, written by the app between turns, produced by tools, or edited by the agent.

Tools are host-provided capabilities. They can query app data, call APIs, or create side effects when the developer exposes them.

Artifacts are durable large files and tool outputs produced inside the harness. The model can inspect them without spending context on every byte.

Persistent Dirs are opt-in developer-backed directories for cross-session files.

Trace is the host-facing record of model calls, tool calls, file activity, artifact writes, commits, and failures.

What To Read Next

• Sessions explains auto-created sessions and continuation.
• Files And Artifacts explains the virtual filesystem.
• Tools Vs Exports explains ordinary tool calls and deliberate app writes.
• Persistent Dirs explains cross-session storage.
• Tracing explains observability.