Agents? What Copilot Is Really About

Earlier this year everyone started calling Copilot an “agent”. Worth straightening out the vocabulary before it ends up in a government procurement doc.

What Copilot was, minus the marketing

Autocomplete with context. A very good one, credit where it’s due: it reads the open file, some of the neighboring files, and predicts the most likely continuation. For boilerplate, repetitive tests and that twenty-case switch, it’s a tendinitis-saving machine.

But predicting the next line is not the same as solving a problem. Copilot in suggestion mode has no goal, verifies nothing, doesn’t know whether its proposal compiles, and forgets everything the moment you close the file. It’s an eager intern with anterograde amnesia: types incredibly fast, remembers nothing about yesterday’s meeting.

What turns a model into an agent

Four things, and all four are engineering, not model:

A goal: a verifiable end state, not a loose prompt. “pnpm test passes” is a goal. “Make this better” is a wish.
Tools: read files, edit, run commands, query the database. Without tools all a model can do is have opinions.
A loop with verification: act, observe the actual result, correct. A model that never sees the output of its own action is guessing with extra steps.
Limits: what it can touch, how much it can spend, when it stops. An agent without limits isn’t autonomous, it’s a pending incident.

Take away the loop and you have expensive autocomplete. Take away the tools and you have a chat. Take away the limits and you have an entertaining story for the postmortem.

Why the demo always works

Because the demo is a brand-new repository, with a self-contained task, no permissions, no real data and no three years of accumulated weird decisions. Any model shines there.

Your repo has a module that only compiles with an undocumented environment variable, a test that fails on Tuesdays, and a convention that lives exclusively in the head of someone who already quit. The agent doesn’t fail from lack of intelligence; it fails from lack of context, and the context is your responsibility, not the model’s.

What did change, and it’s not nothing

A model that can run the test suite and read the error changes the economics of specific tasks: mechanical migrations, mass renames, boring test coverage, translating between APIs. Real work, measurable in hours.

Where it’s still bad: deciding what to build, negotiating an ambiguous requirement, knowing that the elegant solution breaks a contract with another team. Which is to say, exactly what makes a senior a senior.

The boring advice

Before you buy the word “agent”, ask three things: what tools it has, how it verifies it’s done, and what happens when it gets it wrong. If the answer to the third one is “the user reviews it”, you didn’t buy an agent. You bought a review-work generator.

Which can be fine, as long as you know that’s the deal.