---
title: "How Coder works"
description: "Coder sits between you and the coding agents you already use. It makes the quick calls itself, sends only the hard part of the work to the big model, and holds the few steps that are hard to undo…"
url: "https://openagents.com/docs/coder-how-it-works"
---

# How Coder works

Coder sits between you and the coding agents you already use. It makes
the quick calls itself, sends only the hard part of the work to the big
model, and holds the few steps that are hard to undo until you say yes.
This page shows how, with the numbers from our own test.

## Most agents send every step to the big model

![Another harness: three messages run straight into a large glowing circle, the big model. A thin ring around it holds only its own commands: Bash, Read, Edit, Write, Grep, and Web. Every result goes back into the model. Plain Claude Code in our test: $4.00 per accepted fix, 284 seconds median.](/static/docs/coder-harness.svg)

In most coding agents, your message goes straight to one big model, and
so does every step after it. Reading a file, searching, running the tests:
each result goes back into the model, and the model reads the whole
conversation again before the next step, so a long test log is read
again on every step after it. That is where the time and the money go.
The numbers in the picture are plain Claude Code's, from the test below.

## System 1 in front

![Coder: the same messages stop at a ring of quick decisions (route, brief, trim, gate, notes, check) around a smaller big model. Two take a short hop to nearby answers, your code and the commands; only the fix goes on to the model. $1.84 per accepted fix instead of $4.00, 121 seconds median instead of 284.](/static/docs/coder-system-1.svg)

In our test, the same model (Claude Opus 5.5) fixed 21 closed issues
three times each: once as plain Claude Code, and once with a briefing in
front and only the commands it needed. Both got 19 of 63 fixes accepted.
With the briefing, each accepted fix cost $1.84 instead of $4.00, and the
median run took 121 seconds instead of 284.
[How we measured it](https://github.com/OpenAgentsInc/openagents/blob/main/docs/inference/briefed-agent-ab.md).

Coder puts a ring of quick decisions in front of the big model. We call
it System 1: small judgments made in a fraction of a second, before the
big model sees anything. The big model is System 2, kept for the part
that needs it, which is writing the change.

- **Route.** Each new message goes where it's best answered. A question
  Coder can answer from what it knows doesn't wake the big model, and
  unrelated new work starts fresh instead of dragging the old
  conversation along.
- **Brief.** Coder finds the files that matter and hands them over, so
  the agent doesn't spend its first minutes searching.
- **Trim.** Long build and test output reaches the agent as the lines
  that decide the next step. The whole output is saved to a file it can
  open.
- **Gate.** A script the agent hasn't read waits for a read first when
  running it blind looks risky.
- **Notes.** The first time the agent works in a folder, it gets that
  folder's own instructions.
- **Check.** An agent working on an issue that says it's done without
  passing its checks is sent back with what failed.

These come together as the **Token savers** plugin, on by default. You
can turn it off in `/plugins`; see [Coder's plugins](/docs/coder-plugins).

## One Coder in front of every agent

![One Coder in front of every agent: Coder at the center, in a ring marked Auto, with Claude Code, Codex, OpenCode, Grok Build, Devin CLI, and Cursor around it. A light steps from agent to agent, as Shift+Tab does.](/static/docs/coder-agents.svg)

Coder isn't another agent to switch to. It's the same chat in front of
the agents you already have: Claude Code and Codex on your own plans,
OpenCode, Coder's own loop, and others such as Grok Build, Devin CLI, and
Cursor when they're installed here.

A new chat starts on **Auto**, and Coder picks who answers. Press
`Shift+Tab` to choose yourself: Auto, then Claude Code, Codex, OpenCode,
and the other agents found on this computer. Any chat can also hand a
piece of work to another agent and keep going. See
[Choose who answers](/docs/coder-agents).

## Hard-to-undo steps wait for you

![Hard-to-undo steps wait for you: from Coder, reading, editing, running tests, and pushing go straight through. A deploy, a merge, a secret, a payment (in gold), and a DNS change each stop at a hold until you press Y to allow or N to deny.](/static/docs/coder-holds.svg)

Ordinary work goes ahead without asking: reading and editing files,
builds, tests, and pushing commits. Five kinds of step always wait for
you, whichever agent is answering: a production deploy, merging a pull
request, reading or changing a secret, a payment, and a DNS change.

Coder shows the exact command and asks one plain question. `Y` allows it
this once; `N` or `Esc` denies it, and the agent carries on with the rest
of the work. See [Safety](/docs/coder-safety).

Next: [Use Coder in the terminal](/docs/coder-terminal).