# Several agents on several harnesses

At the end, you know how a step picks its coding tool, model and thinking level, how a workflow mixes them, and where you sign in.

## At a glance

**Time:** 5 minutes to read, 5 more to sign in to a second tool.

**What you need:**

- At least one of Claude Code, Codex or pi signed in on Settings → Engines

## How it works

A **harness** is the coding tool a step runs in: Claude Code, Codex or pi. An **engine** is a harness, a model it lists and a thinking level that model offers. One engine menu picks all three everywhere: a step in the builder, a step for one run, an agent in the composer, the copilot in its chip. The thinking ladder is off, minimal, low, medium, high, xhigh, max, ultra; each harness says which levels each model offers, and a level a model does not offer is refused, never downgraded.

Each tool signs in with your own subscription, on the machine Loopgate runs on, through its own sign-in: Codex and Claude Code print an address and a code, pi opens its own screen. Loopgate never touches those credentials and never holds an API key for you; an API key is an optional alternative in the environment.

A workflow may put each step on a different engine, and the built-in Design, build, panel does: one of its seven reviewers runs on a second model so the diff is read by two minds. A step nobody chose an engine for gets the first ready harness in the order Codex, Claude Code, pi, on its default model, thinking at high when its persona judges and the model offers it. A workflow that names an engine this machine cannot run is fitted, step by step, to one it can, and the draft records each change.

Every ready harness is also an agent of its own in the Who picker, a plain session with no persona: the tool as it is, with Loopgate's visibility and record around it.

## Why two harnesses

Each model reads what the other missed. Parallel plans, Pair programming and Root cause, fix put two harnesses on the same problem, and a lead or a navigator answers from both: [The patterns](patterns.md). A workflow that names a harness this machine lacks is fitted to one it has, and the draft records each change.

## Try it

1. **On Settings → Engines, press Sign in on a second tool** and follow its address and code.

   You should see: the row saying Ready, and its models listed with their levels.

2. **On Tasks, pick Plan, build, review and click the reviewer's station.**

   You should see: its engine chip; open it and pick the second tool's model and a level it offers.

   You should see: a blue dot on the station and one line, `1 chosen here`, on the card.

3. **Start, and read the reviewer's turn.**

   You should see: its engine named on its reply's chrome, and its cost counted into the run's.

## Next

Change engines, rounds and budgets while it runs: [Steer a run](steer.md).