The Fourth-Agent Test

The first agent is usually a demo. The second starts to show a pattern. By the fourth, you find out whether you have a system or just four scripts with access to the same accounts.

I call this the fourth-agent test. If adding another agent means figuring out permissions, logs, errors, and ownership all over again, the setup is not ready to grow.

Five questions to answer

Every agent should have a clear answer to these questions:

1. What information can it read?
2. Which tools can it use?
3. Which actions require approval?
4. How will the result be checked?
5. What record will be left after the run?

These questions turn a loose prompt into a piece of software you can reason about.

Limit access to the job

A local agent rarely needs access to everything on your computer. It needs the files and tools required for the current task.

I break that access into four parts:

working folder → allowed tools → approval rules → final checks

The working folder controls what the agent can see. The tool list controls what it can do. Approval rules cover actions that are hard to undo. Final checks define what “done” means.

If those rules live only in your head, you have to remember them every time the agent runs. That gets harder with each new agent.

Check the result separately

Agents often check their work using the same assumptions they used to make it. That is convenient, but it can miss the same mistake twice.

For code, run the tests in a fresh process. For a generated file, open and parse the file after it is written. For a change to an outside service, read the new state back from that service.

The rule is simple: the agent saying “done” is not proof that the job is done.

What should exist by agent four

By the time you add a fourth agent, you should have:

• one place to define tool access;
• a shared way to look, change, test, and report;
• readable activity logs;
• clear approval rules;
• a way to compare the expected result with the real one.

If you do not have those pieces, build them before adding another agent. They will save more time than agent five will.

A final check

Before using an agent regularly, ask:

Can someone tell what this agent was allowed to do, what it actually did, and how we checked the result?

If not, it is still a demo.