LoopgateTry it

Docs / Workflows / A model per step

A model per step

MarkdownOpen in ClaudeOpen in ChatGPT

At the end, one workflow runs its steps on two coding agents, each step on the model and thinking level you chose for it, and you have changed one while it ran.

Time

10 minutes.

You need

Steps

  1. On Settings → Engines, sign in to a second agent. Press Sign in on its row and follow the address and code it shows.

    You should see: the row says Ready, and its models button lists each model with the levels it offers.

  2. On Tasks, pick plan-build-review under Who and click the review step under the box.

    You should see: its engine, access, pace, persona and instructions for this run.

  3. Open its engine and pick the second agent, a model and a level.

    You should see: a dot on the review step and 1 step changed on the line above it.

  4. Say a task and press Run.

    You should see: the review step's engine on its reply in the conversation, and its cost counted into the run's.

  5. While it runs, open Setup in the run and change the build step's engine to another agent.

    You should see: next turn · new conversation under the row, and one line in the conversation naming the change and when it lands.

Good to know

  • One engine menu picks the agent, the model and the thinking level everywhere: a step in the builder, a step for one run, an agent on Tasks and the copilot.
  • The levels run off, minimal, low, medium, high, xhigh, max, ultra. Each agent says which levels each model offers. A level a model does not offer is refused, never lowered.
  • A step nobody chose an engine for gets the first agent that is Ready, in the order Codex, Claude Code, pi, on its default model, thinking at high when the step's 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. The run records each fitted engine as a change at Start.
  • A change of agent starts a new conversation for that step at its next turn. A change of model on the same agent keeps the conversation when that agent can switch models mid-session.
  • pi offers only the models of the providers it is signed in to.
  • Every agent signs in as you, on the machine Loopgate runs on. Loopgate never holds the credential and needs no API key.
  • Each Ready agent is also in the Who picker as a plain session: the agent as it is, with no persona.

If something is off

  • A model is missing from the menu. Its agent is not Ready, or, for pi, its provider is not signed in. Sign in, then press Detect again on Settings → Engines.
  • The level you want is not listed for a model. The model does not offer it. Pick another model that does.

Next