HarnessGrad

HarnessGrad: Harness as A Continual Learner

Keep the model fixed. Let the harness evolve.

Tools, context, and control flow shape what an agent can do. HarnessGrad turns environment feedback into experiments on that editable code.

Start an experiment
An agent is a harness plus a model, working within an environment.An original little agent sits inside a teal adjustable scaffold. The model stays fixed while tools, context and control flow can change. Actions go to the environment and feedback returns.Environmenttools · context · control flowmodelactionsfeedback
Agent = Harness + ModelA fixed model. A changeable way of working.

Five ways to learn.
One overlooked learner.

An agent can learn in its context, memories, skills, or weights. It can also change the harness that puts them to work.

HarnessGrad focuses on this last path: improving how an agent organizes tools, builds context, and acts in its environment.

These paths can work together. Their cadences overlap; this is a map of what changes, not a speed ranking.

  1. Context

    Within an interaction

    A few words change the next step.

  2. Memory

    Across tasks

    Experience leaves a persistent trace.

  3. Skills

    As experience accumulates

    Useful procedures become reusable.

  4. Weights

    During training updates

    Learning changes model parameters.

  5. Harness

    Between experiment rounds

    Feedback reshapes how the agent works.

Different learning surfaces. The harness is ours to explore.

Your harness.
Its next experiment.

Keep your harness’s execution logic. Connect it through an adapter, choose a supported environment, and declare what can change.

An improvement method proposes an edit. HarnessGrad runs the candidate, gathers feedback, and evaluates what happened.

Read the integration guide
One little agent, an evolving harnessIllustrative experiment, not measured results. The same robot and model remain fixed. A run sends an action, feedback returns as a trace, the harness gains an edited tool shelf, and evaluation compares the original with the candidate.Environmenttools · context · control flowmodel stays fixedactionrun tracecontexttoolsoriginalcandidate
Run → Feedback → Edit harness → EvaluateIllustrative cycle, not measured scores.
Same model throughout; no gain is assumed.
  1. Run

    Give the agent a task.

    Your entrypoint starts a run. A model and environment bridge connects reasoning to actions, using the tools and context your harness provides.

  2. Feedback

    Bring the experience back.

    Collect environment feedback and the run trace. An improvement method examines this experience to decide what to try next.

  3. Edit harness

    Change the scaffold, not the model.

    The method proposes an edit within your declared editable surface: tools, context construction, or control flow. The model stays fixed.

  4. Evaluate

    Test the change, then decide.

    Run the candidate under declared evaluation conditions. Compare the original and candidate evidence, then use what you learn to plan another round.

Let the evidence speak.

Explore recorded runs and their conditions, or start a small experiment of your own.