gezel Gezel Help

Generalist mode

A task in gezel is a plan walked one step at a time, with a check at the end of each step. There are two ways to staff that walk.

Stepwise is the crew approach: each step names the trade it needs, a specialist for that trade picks it up in a fresh conversation, passes the step's check, and hands the next step to the next specialist. It is how small on-device models get the most out of a plan, because each conversation stays short and focused.

Generalist mode puts one pair of hands on the whole task. One owner — the Generalist — reads the outline of the task, works the active step, passes its check, and moves to the next, all in one continuous conversation. What it learned in step two is still in front of it in step five. The steps and their checks are the same ones a crew would walk; nothing is skipped and nothing is merged.

When it is on

Settings → Artificial intelligence → Run in generalist mode has three keys:

  • Automatic (the default): when the recipe states its goal, one gezel works from that goal (see below). Otherwise generalist mode is on for hosted frontier models — Claude, ChatGPT and Codex, GitHub Copilot — and off for models running on this device.

  • On: every task runs in generalist mode, whatever the model. Useful for trying a capable local model on a whole task.

  • Off: every task runs stepwise, even on a frontier model.

The choice is made once, when a task is created, so changing the setting never disturbs a task that is already under way.

Working from the goal

Some recipes state their goal outright: what to deliver, and the bar it has to clear. For those, Automatic hands one gezel the goal, the files to deliver and the checks they must pass, and lets it decide the steps itself. The checks run on the finished work instead of after each step. In our tests this did about as well as walking the recipe step by step, in well under half the time. A few models we measured also get the recipe alongside, as advice they may follow or skip. Choosing On or Off brings back the step-by-step walk.

What stays the same

  • Every step, every check. The Generalist advances one step at a time and each step's check must pass before the next is revealed.

  • Fan-out still fans out. When a step spawns one child task per item — one story per topic, one invoice per client — each child is its own conversation, so the work still happens in parallel. The Generalist collects the results when the children finish.

  • Your choices win. If you assigned a task to someone in particular, they stay the owner; a step that is yours to approve stays yours.

What changes

  • One owner, one conversation. No hand-offs between specialists, no re-reading of notes to reconstruct what the previous step found.

  • The whole outline in view. The owner sees the goal and every step, marked done, active or pending, and can shape today's step for tomorrow's.

  • The full bench. The tools every step in the plan needs are available from the start.

See also Tasks and supervision for how steps, checks and pauses work.

Watch this article as a slideshow · Read it as markdown, for AI agents