Building an agent is easy. Operating one is not

The afternoon and the months after

Building an LLM agent in n8n is quick: a trigger, a model node, a few tools, and it works on the first try. That afternoon is the part every tutorial covers.

What comes next is not covered, and it is where the incidents are. Is the agent healthy right now? What did it cost this week? What exactly did it do at 3am last Tuesday? How do you stop it when it starts looping on a task it cannot finish? A demo does not need answers to these questions. A system that runs unattended does.

Four things a demo does not have

Admission control. Before an agent starts a run, something should check that it is allowed to: a budget that has not been spent, a limit on how many runs happen at once, and a kill switch an operator can flip without editing the workflow. Without it, the first runaway loop is stopped by the invoice.

A plan you can check. An agent that decides its next action in free-form prose is hard to audit and easy to derail. Asking for structured steps first makes each run inspectable before it acts.

A verification step that ends the loop. After each action, a separate check asks whether the step did what the plan said. This is what makes the loop terminate: an agent that only acts never knows it is done, so it keeps going.

Observability you will actually read. Log the run, the steps, the tools called and the cost, in a form you can search by run. Alert on a few signals that mean something, such as budget burn or repeated failures, and deliberately ignore the rest, because an alert channel nobody reads is the same as none.

Rehearse the bad night

Write down, before it happens, what you do in the first five minutes of the likely incidents: the agent loops, the budget is exhausted, a tool starts failing, the model starts refusing. A short runbook turns a 3am page into a checklist instead of an investigation.

Putting it into practice

The n8n Agent Operations Suite provides two workflows for this: an orchestrator that runs plan, act and verify with budgets, a tool allowlist and a kill switch, and a scheduled health check that runs a canary and reports on budget. It adds the planning and verification prompts, a runbook of seven incidents with their first five minutes, and observability guidance on what to log and alert on.

Measuring whether a change to an agent setup actually helped is covered in "it felt better" is not an agent benchmark.