# Start with the task, not the settings

> Grain’s new task composer removes setup from the critical path, adds one shared model picker, and brings GPT-6 Astra to the work that needs it.

- Canonical page: [/blog/start-with-the-task-not-the-settings](https://grain.sh/blog/start-with-the-task-not-the-settings)
- Agent-readable page: [/agents/blog--start-with-the-task-not-the-settings.md](https://grain.sh/agents/blog--start-with-the-task-not-the-settings.md)
- Section: Blog
- Category: Product
- Published: September 4, 2026
- Reading time: 6 min read

## A task should begin with the work

Starting an AI task had quietly become a configuration exercise. Before a teammate could describe what needed to change, they were asked to think about a model, a machine, a repository, a policy, and an execution path. Those controls were useful, but putting every decision first made the simplest tasks feel heavier than they were.

The new Grain composer starts with the brief. Write what you want done, attach the files that explain it, and start. Grain chooses a sensible execution path from the capacity already connected to the workspace. The deeper controls are still there when the work genuinely needs them; they no longer stand between an idea and its first useful run.

## Less setup, without losing control

The lighter flow is not a smaller product. It is a better order of operations. Grain can default to hosted Grain Models or an available connected computer, while the composer keeps model, repository, and policy choices close at hand. Teams that know exactly how a task should run can still specify every part. Everyone else can let the system make the routine decisions.

Drafts now persist with their attachments and setup choices, too. Closing the composer no longer means reconstructing a half-written brief later. If a task touches several repositories, you can select them together and Grain gives each one an isolated task branch and working copy. Finished artifacts can sync to a new GitHub or Bitbucket branch without merging into the codebase.

- Begin with a brief and attachments, then reveal setup only when it helps.
- Let Grain choose hosted or connected-computer capacity by default.
- Keep drafts, files, and execution choices when you leave the composer.
- Work across multiple repositories without touching an active checkout.

## One model picker everywhere

Model choice is now consistent across task creation, task chat, AI teammate threads, and workrooms. Grain Models and the models available on your connected computers appear in the same picker. Online computers come first, unavailable options explain why they cannot run, and choosing an Executor model selects the computer capable of running it.

The selection persists into the messages that follow, so the conversation does not silently jump between execution paths. A chosen Executor model takes precedence over a workspace direct-API chat key, while Grain Models keeps Auto routing for teams that want Grain to place the work.

## GPT-6 Astra joins the roster

GPT-6 Astra is now available anywhere a Codex-backed Executor model can be selected: in the main picker, provider lists, and per-agent overrides. That makes it possible to use Astra for one demanding task or assign it to a specialist teammate without changing the workspace default.

The bigger change is not one additional model. It is a clearer way to decide. Start with the work, choose a model only when the choice matters, and keep that choice consistent wherever the task continues.

## About Grain

Grain is a multi-agent work platform. A person describes an outcome as a Mission, Grain coordinates the models, tools, files, repositories, and connected computers needed to do the work, and the result returns for review.

Start with [the complete agent-readable index](https://grain.sh/agents/index.md) or read [the full site corpus](https://grain.sh/llms-full.txt).