A coding agent, and the difference from an autocomplete

The difference between a coding agent and an autocomplete is that an agent acts and then observes, so its usefulness depends almost entirely on whether the things it can observe actually tell it the truth about what it did.

A coding agent runs a loop: read the code, make a change, run something, look at the result, keep going. That is a different thing from a model that suggests the next line.

Acting is not the interesting part

Editing files and running commands is mechanical, and every agent does it competently. What separates a useful agent from a frustrating one is the quality of the feedback it gets back after each action, which is a property of your project rather than of the agent.

An agent is only as good as your test suite

It changes something, runs the tests, and believes them. A thorough suite makes an agent effective, because a mistake is caught in the same loop that made it. A thin suite makes it dangerous in a specific way: the agent gets a green result, concludes it succeeded, and builds on the mistake.

The check that cannot fail is the worst input you can give it

A test that passes regardless tells the agent everything is fine, forever. It will keep going, keep building, and be confidently wrong the entire time. Weak checks are worse than missing ones here, because a missing check leaves the agent uncertain while a broken one makes it sure.

What to give an agent that it cannot fake

Something outside itself. A type checker, a real test against a real system, a live source that will contradict it. An agent working entirely within a world it can also edit has no way to be wrong, which is precisely why it will confidently produce something that does not work.

Reviewing what an agent did

Read the parts that handle money, failure and staleness. Those are where a plausible change and a correct change diverge most, and they are the places an agent's own checks are least likely to have covered. The rest can usually be judged by whether it works.

Live, right now, on this page

Market	Price	Funding	24h volume
BTC	$81,754.50	0.0013%	$180,448,812
ETH	$2,413.35	0.0013%	$154,879,987
SOL	$103.67	0.0013%	$34,236,995
HYPE	$78.76	0.0013%	$6,599,165

Read from a live market as this page rendered, through a public endpoint that needs no key. Read at 2026-10-18 06:45 UTC; accurate as of that time and not afterwards.

Describe something and watch it get built.

Open the builder, or start from a working app and change it.

Common questions

What is the difference between a coding agent and autocomplete?

An autocomplete suggests the next line. A coding agent runs a loop: it reads the code, makes a change, runs something, looks at the result and keeps going, so its usefulness depends on whether what it observes tells the truth.

Why does a coding agent need a good test suite?

It changes something, runs the tests and believes them. A thorough suite catches a mistake in the same loop that made it. A thin suite gives the agent a green result, so it concludes it succeeded and builds on the mistake.

What should I review after a coding agent works?

The parts that handle money, failure and staleness. Those are where a plausible change and a correct change diverge most, and where an agent's own checks are least likely to have looked.