An agent is more than its prompt

A strong prompt can make an agent look alive for one run. A production system has to survive the next hundred runs: stale documents, broken tools, partial access, repeated requests, bad inputs, and tasks where the correct answer is to stop. That requires a harness around the model, not just better wording inside the model.

The harness decides what context is available, which tools exist, how sessions are separated, how writes are confirmed, how errors are shown, and how results are evaluated. It is the operating environment. Without it, an agent is only a fluent interface over undefined behaviour.

The best first production agents own a clear slice of work. They review documents, prepare approval packets, search a controlled knowledge base, reconcile records, or coordinate a specific development task. A vague employee-shaped agent is harder to test, harder to trust, and harder to improve.

Once the job is named, evaluation becomes real. Measure source quality, tool success, escalation rate, latency, user edits, cost, and failure modes. The point is not to remove humans from judgment. The point is to move repetitive work into a system that shows enough evidence for humans to intervene at the right level.

Agents need logs, fixtures, regression tests, prompt or policy versions, and a place to inspect failures. Tool schemas change. Documents drift. Users discover new edge cases. If no one owns the loop, the agent becomes an impressive demo that slowly stops matching reality.

This is why iftodo treats agent work as software work. Build the connector, name the authority boundary, capture the evidence, test the unhappy path, and keep the system readable. A serious agent is not more magical than normal software. It is normal software with a model inside the loop and a sharper need for traceability.