What Is an Agent Harness? The Scaffold That Turns an LLM Into a System
An agent harness is the scaffolding around an LLM that makes it reliable — tools, context, loops, and guardrails. The anatomy of a well-performing agent.
A harness is everything between the model call and the outcome: the tool surface it can reach, the context it is handed, the loop that decides whether to call again, and the verifier that decides whether it is finished. Change the model and the numbers move a little. Change the harness and the numbers move a lot. When comparing runtimes, use this anatomy rather than a feature checklist: evaluate the framework’s harness surface for explicit state, authority, control flow, evidence, and release gates.
The four organs
Tools, context, control flow, guardrails. Each one fails in a characteristic way, and each one has an instrument that tells you it is failing before your users do.
01 · Tools
Few, sharp, validated
A small surface with strict schemas beats a broad one with permissive types. Every argument the model can get wrong is a failure mode you have to eval.
02 · Context
Window hygiene
Retrieval discipline and compaction decide what the model knows at inference time. Prompts only reorder what is already there.
03 · Loop
Stop conditions
Most agent failures are loop failures: no verifier, no retry budget, no honest definition of done.
"Model benchmarks do not predict agent performance. Eval the harness."
The rest of this series takes one organ at a time and runs the experiment: a fixed task suite, a single variable, and the traces published alongside the result.