Skip to main content
Features

Better Harness

Review the engineering setup around agent work and turn findings into repair plans.

Better Harness reviews the environment around an agent—not just the answer from one task. It helps you find gaps that repeatedly make agent work harder to understand, execute, validate, or deliver, and turn useful lessons into maintainable project assets.

What is an Agent Harness?

An agent usually works in a loop: understand the task, take action, inspect the result, and adjust. A reliable loop requires more than a capable model. The agent also needs clear project context, usable tools, operating boundaries, validation methods, and a way to retain lessons from earlier work. Together, these supporting mechanisms form the Agent Harness. Depending on the project, a harness may include repository instructions, rules, skills, hooks, plugins, connectors, scripts, test commands, release checks, and human review steps. Its purpose is to help the agent answer four practical questions:
  • What outcome is expected, and what is out of scope?
  • How should the project be operated and changed?
  • What evidence proves that the result is correct?
  • What should happen when an operation or validation fails?

Why use Better Harness?

Harness gaps are often hard to notice in a single successful task. They appear over time: the same instruction has to be repeated, project conventions remain undocumented, validation is skipped, or feedback from a review never reaches the next task. Better Harness examines these recurring patterns and helps identify the smallest durable improvement that can prevent them from happening again.

What it reviews

A useful harness gives the agent enough structure to complete work and prove the result. The review can examine areas such as:
  • task goals, scope, and project context;
  • reusable agent assets, including rules, skills, hooks, plugins, and connectors;
  • execution entry points, controls, and repeated manual work;
  • validation evidence, delivery steps, and session outcomes;
  • project feedback that should become a durable rule, workflow, or tool.
The purpose is to identify a concrete gap and the smallest maintainable improvement—not to add configuration for its own sake. For example, a missing project convention may belong in a rule, a repeatable procedure may be better expressed as a skill, and a check that must always run may be better implemented as a hook or script.

How a review leads to improvement

A Better Harness review follows a practical loop:
  1. Map the current harness — identify the task context, execution path, reusable assets, validation steps, delivery process, and available feedback.
  2. Find the breakpoint — locate where the workflow lacks clear instructions, an actual integration, execution evidence, or a recovery path.
  3. Choose the smallest durable fix — match the problem to an appropriate rule, skill, hook, script, workflow, or human review step.
  4. Verify the improvement — apply the change in a reviewable task, run the relevant checks, and confirm that a representative future task uses the new mechanism.

Start a review

  1. Open a task in the workspace you want to review.
  2. Type /better-harness in the task input.
  3. Select better-harness:better-harness from the slash-command suggestions.
  4. Add the scope or problem you care about, such as repeated release mistakes, missing test evidence, or project instructions that agents regularly overlook.
  5. Send the task and let the review inspect the current project and relevant task history.
Start with a specific symptom. “Why do release tasks repeatedly miss link validation?” produces a more actionable review than “improve everything.”

Review the findings

For each finding, check:
  • the evidence that shows the problem exists;
  • the part of the workflow affected;
  • the proposed durable fix, such as a rule, skill, hook, script, or human approval step;
  • how the improvement will be verified in a later task.
Treat the report as a decision aid. Confirm recommendations against the repository and your team process before starting a repair.

Apply improvements safely

  1. Choose one high-value finding with a clear acceptance criterion.
  2. Start a separate task for the repair so its file changes and tool actions remain reviewable.
  3. Run the project-specific checks named in the plan.
  4. Re-run Better Harness or a representative task to verify that the new control is actually used.
Do not turn transient failures or one-off preferences into permanent project rules. Durable assets should solve a repeated problem and remain understandable to future maintainers.