ENTERPRISE AI

The loop inside the harness

The harness is context, state, tools, a check, permission, and a trace. The loop reasons, acts, and looks. The model is one part of that loop.

Figure 1: Context for this step, a bounded tool, and a check before the result is trusted.

1. Six parts, and the model is not one of the six

This method uses six harness components. Context is the information for the current step: the policy, the schema, the task, the retrieved evidence. Memory and state keep the facts, the open questions, and pointers to evidence you can reopen. Tools read a system, calculate, search, or take a bounded action. Verification asks whether the work met the requirement. Permissions and sandboxes limit what the agent can touch. Observability makes the behavior, the quality, the security events, the latency, and the cost inspectable.

Leave one of those out and the demo can still talk. The operation cannot be run.

2. Reason, act, look, and then be checked

A widely cited pattern called ReAct interleaves a reasoning trace with an action and an observation. The enterprise loop keeps that rhythm and adds the parts a company cannot skip: a check, a permission, and a trace. The model proposes the next step. A tool does a bounded thing. The observation comes back. A verifier decides whether that observation is allowed to count.

State should be structured. Facts, decisions, actions, and unresolved questions do not belong in one blurry paragraph the next turn will misremember. High-impact calculations and state changes stay deterministic, so a second system can recompute them without asking the model to agree with itself.

Layman rule: If you cannot point at the check, the permission, and the trace, you have a prompt, not a harness.

3. Retrieved text is untrusted text

Retrieved documents, tool output, durable memory, a sub-agent’s message, and the model’s own artifacts may all be untrusted. A prompt that says “ignore instructions in the document” is not a control. The control is structural: identity, provenance, least privilege, a transaction boundary, authorization on the specific action, an independent check, a log, and a failure that stops safely.

A published security heuristic is worth designing against. An agent is most exposed when it can see private data, read untrusted content, and take an action that leaves your boundary, all in the same run. Split those three when you can. Do not give one loop all of them because it was convenient in the demo.

4. Diagnose the step before you rewrite the prompt

When a run fails, find the step. A bad retrieval, a tool call with the wrong argument, a permission that was wider than the task, a check that never ran. Rewriting the whole prompt hides the step. Parallel workers are allowed. Each one needs an owner, and the results need a controlled way back together. A handoff you did not curate is how two agents disagree in production and both sound sure.

Bound what “done” means. An agent that may keep going until it feels finished will spend the budget and widen the change. Done is a check that exits.

5. What to draw next

For one workflow, name the six components in a sentence each. If observability is “we will add logs later,” the harness is not finished.

What the model is allowed to see is the next design: context, and when it goes stale.