Harness
also called Scaffold, Agent framework
Everything around the model that makes it an agent.
The software wrapper that gives the model its tools, instructions, permissions and context management, and that runs the agent loop. The model reasons; the harness does everything physical.
Two products using the identical model can differ enormously in usefulness purely because of harness quality — how context is assembled, how tools are described, how failures are surfaced. It is where most engineering effort lives.
Harness quality is where much of the perceived difference between products lives, and it is also why model-versus-model comparisons so often mislead: the same model behind a better harness outperforms it behind a worse one. When you evaluate, evaluate the pair.
Swapping harnesses beat swapping models
Two teams use the same model; one reports far better results on the same task.
The difference is tooling: a real filesystem, a code-execution tool, tests wired in, context managed properly. The lesson is to improve the harness before reaching for a more expensive model.
What the harness owns
Listing the components you are responsible for.
Tool definitions and execution, permission checks, context assembly and trimming, retries, logging, cost accounting and the exit condition. The model only supplies judgement; everything else is yours to get right or wrong.
- blogDemystifying evals for AI agents ↗
Gives working definitions for harness, task, trial, grader, transcript and outcome, and argues you must evaluate the model and harness together and judge final environment state, not just the last message.
- blogEffective harnesses for long-running agents ↗
The initializer-plus-incremental-sessions pattern for work that outlives a context window, and the four ways it fails: doing too much, losing context between sessions, declaring victory early, not testing end to end.
- blogScaling managed agents: decoupling the brain from the hands ↗
Separates model, harness, durable session log and execution sandbox — the architecture that makes long-running agents survive crashes and tool boundaries.
Agent
A model given tools, a goal and permission to keep going on its own.
Agent Loop · Agentic loop
Think, act, observe — repeated until the task is finished.
Context Engineering
Deciding what goes in the window, in what order, at all.
Tool
A function the model is allowed to call to affect the world.
Subagent
A helper agent with its own clean context window.