GLOSSARY

What is an agent harness?

An agent harness is everything between the model and the task: prompt assembly, tool definitions, the loop that picks the next action, the environment where actions land, and the rule that ends the run. Swap the harness and the same model behaves differently. Harness quality, not just model choice, predicts whether a task finishes.

The model is one component. The harness decides what it can see, what it can do, and when the work counts as finished.

Start free See pricing

An agent harness is the software around a language model that gives it tools, a loop, and a place to act, turning a single completion into a multi-step task.

The five parts

A harness has five working parts. Prompt assembly decides what the model sees at each step, including tool results and history. Tool definitions describe what it can call and how. The loop runs those calls, feeds results back, and decides whether to continue. The environment is where the calls take effect: files, a shell, a browser. The stop rule decides when a run is done, or blocked.

Harness, model, framework

These terms get used loosely, so keep them apart. The model is the reasoning engine. The harness is the runtime that turns reasoning into actions. A framework is a library you use to build a harness. Two teams can run the same model in different harnesses and get different results, because one feeds back test output and the other does not. METR reports agent results per scaffold, not per model alone.

What good looks like

A good harness makes the situation legible. Tool results come back trimmed but complete enough to act on, failures are reported as failures, and the next step is chosen from evidence rather than from the model’s memory of what it intended. It also bounds itself: a step budget, a time budget, and a clear stop when the work is verified or genuinely blocked.

How to evaluate one

Give it a task with a checkable outcome and watch three things. Does it find the right files without being told? When a command fails, does the next step use the error, or repeat the same call? And when it stops, can it show why it believes the work is correct? A harness that answers those well is worth more than a stronger model behind a weaker loop.

How Hyperlane implements it

The Hyperlane harness is the agent loop: plan, execute tool calls in dependency-aware parallel waves, verify, repair, deliver. It runs inside a per-run sandbox, so tool calls land on a real machine rather than a simulated one. A verify-before-done gate blocks completion after edits until commands have passed. The same loop backs the workspace, API, SDK, CLI, and MCP server.

What Hyperlane ships

  • Autonomous agent loop that plans, edits, runs, and verifies work end-to-end
  • Every run executes inside an isolated, resource-capped sandbox
  • Live execution traces - diffs, terminal output, and reasoning as it happens
  • GitHub-native workflow: pushes branches and opens pull requests itself
  • Web app, REST API, TypeScript SDK, CLI, and MCP server interfaces
  • Dollar-for-dollar credits - pay $20, get $20 of agent work

Frequently asked

Is an agent harness the same as an agent framework?

No. A framework is a library for building agents. A harness is the assembled runtime a specific agent runs in, including its prompts, tools, environment, and stop conditions. You can build a harness with a framework, or without one.

Why does the harness matter if the model does the thinking?

Because thinking alone does not change a repository. The harness decides which files the model can read, what it learns from a failed command, and whether its claim of success is checked. Those choices change outcomes.

Does the harness include the sandbox?

Usually yes, in the sense that the harness needs somewhere for tool calls to land. The sandbox is the isolation boundary; the harness is the controller. Keeping them separate in your head helps, because you can change one without changing the other.

Can a different harness run on the same platform?

In principle yes, since the harness sits between the platform and the model. On Hyperlane, which harness drives a run is resolved by the platform and applied per launch, so a session does not split across two loops. Changing it is an operator action, not a per-user setting.

What should a harness never do?

Report success it cannot show. A harness that accepts the model’s summary as proof of completion teaches the model that summaries are the deliverable. The stop rule should depend on evidence the harness collected itself, such as an exit code or a passing test.

Last verified 11 Nov 2026.

Sources: Anthropic - Building effective agents METR - Evaluating frontier AI R&D capabilities Hyperlane docs - How a run works

Keep exploring

Delegate the ticket. Review the PR.

Start with the Free plan — 10 minutes of agent computer time every month, no card required.

Try Hyperlane free

Hyperlane AI is not affiliated with the Hyperlane blockchain protocol (hyperlane.xyz) or the Hyperlane code editor (hyperlaneide.com).