The Agent Harness Is the Product

A frontier model can write code, inspect logs, critique a design, and decompose an unfamiliar task. Yet the same model can be brilliant in one environment and strangely ineffective in another.

The difference is often the harness.

By harness, I mean the system that makes a model operational: context, skills, tools, execution, memory, policy, observability, verification, and the loop that decides what happens next.

Intelligence needs a world#

A model arrives with broad prior knowledge. It does not arrive with a reliable map of your organization.

It does not automatically know which dashboard is authoritative, which runbook is stale, which service owns a dependency, which apparent error is normal during a migration, or how much risk this rollout may accept. It may know twenty reasonable ways to review a deployment. It does not know which one your organization considers complete.

That missing context is not a prompt-writing inconvenience. It is product infrastructure.

The harness answers six practical questions:

1. What can the agent perceive?
2. What can it do?
3. What rules constrain it?
4. What state survives between steps?
5. How does it know that the work is good?
6. What happens when it is uncertain or wrong?

If any answer is vague, model capability leaks away as confusion, wasted exploration, or unsafe confidence.

Skills are navigation, not the whole runtime#

Skills are a powerful abstraction when they package a repeatable way of working: when to use a workflow, what evidence it needs, how to select tools, and how to verify completion.

But a skill is not a substitute for permissions, scheduling, isolation, or policy enforcement. Telling an agent how to inspect production does not decide whether it is permitted to mutate production. Describing a checkpoint loop does not create a durable timer. Writing “keep customer data private” does not create an access boundary.

Good harnesses separate these concerns:

• Skills encode reusable procedure and judgment.
• Tools expose typed capabilities.
• Policies bound authority.
• Sandboxes isolate execution.
• State machines make long-running work durable.
• Evaluators test outcomes and behavior.
• Memory carries forward useful precedent.

The model can compose them dynamically. Their responsibilities should remain distinct.

Progressive disclosure beats encyclopedic context#

The natural response to an underspecified environment is to write a larger instruction file. This works—until it does not.

As guidance grows, old rules linger, important constraints compete with trivia, and the model spends attention on material irrelevant to the current step. When everything is always loaded, importance loses meaning.

A better harness exposes a small, stable map and lets the agent retrieve detail when it becomes relevant. This is the same reason good software has interfaces rather than one global namespace.

Context should be addressable, scoped, fresh, and attributable. The system should know not only what a document says but why it was selected and whether a stronger source supersedes it.

Dynamic orchestration, static accountability#

Agents benefit from being able to form a plan, create focused sub-agents, and choose tools based on the work they discover. Hard-coding every future workflow throws away much of their value.

But dynamic execution does not require dynamic accountability.

The system can allow an agent to invent a strategy while keeping stable records of delegated goals, granted capabilities, evidence gathered, state transitions, costs, and final authority. Strategy may emerge at runtime; accountability should not.

This is the balance I want from a universal harness:

Let the model construct the path. Make the system preserve the boundary and the proof.

The durable moat#

Models will continue to improve, and generic orchestration will become easier. That does not make the harness irrelevant. It changes where differentiated value lives.

The durable advantage is not a clever prompt or a fixed graph of agent calls. It is the accumulated, operational representation of a real environment:

• topology and ownership that stay current,
• tools with safe and useful semantics,
• policies tied to actual authority,
• historical outcomes connected to decisions,
• evaluators calibrated to the domain,
• and workflows that reveal the right context at the right time.

That infrastructure lets each new model become useful faster. It also survives when the model changes.

The model is extraordinary leverage. The harness is how an organization converts that leverage into work it can depend on.