Agent Harnesses Explained
What sits between a model and the world.
An interactive guide to the agent harness: the loop, the tools, permissions and context management around a model, animated by a tested simulator.
- turn 1505
- turn 2561
- turn 3689
- turn 4777
- turn 5871
- turn 6939
- System prompt
- Tool definitions
- Task
- Reply header
- Model's messages
- Tool results
How this is checkedAgent_Loop_Sim's Python reference and the TypeScript port in your browser produce the same events, token for token and float for float, and three recorded runs of a small open-weights model are replayed token for token. How it is tested →
What you'll learn
A chat model only writes text. The harness around it turns that into an agent: it runs the loop, describes the tools, parses the calls, decides which ones may run, keeps the context inside the window and the prompt cache warm. To fix one failing test, a scripted agent here makes 6 model calls and sends the model 4,342 tokens to write 197 tokens.
Each chapter takes one of the harness's choices apart, built around an animation. Every frame comes from Agent_Loop_Sim, a deterministic simulator with a real tokenizer, tested against its Python reference. Nothing here calls a live model: the scripted runs are reproducible, and three runs of a small open-weights model were recorded once and are replayed token for token, failures and all.
Part of a family of sixteen companion sites in four tracks, all in the site menu above: Foundations (Decoder, Training), Hardware (Kernels, Systolic, Accelerators, Racks), Inference (Serving, Architectures, Numerics, Trade-offs) and Agents (Harnesses, Protocols, Context, Orchestration, Evals, Security). It is the first of the six agent sites. How this one was built, and how to check it: about.