Multiple AI assistants: dividing tasks and checking results

What AI can coordinate

For its redesigned projects in Claude Code, introduced on 17 September, Anthropic describes a coordinating session that assigns subtasks and brings the results together. The work runs on the provider’s servers and can continue when you step away from your computer. Claude Code is mainly known as a software development tool.

Meta’s Muse is a personal assistant also intended to work on tasks for longer periods. The two announcements emphasise different things; working in parallel and working in the background are compatible. Details and the status at the time of each announcement are in the linked briefing.

Read about both announcements in the AI Briefing ↗

Source: Anthropic: Projects redesigned · Meta: Introducing Muse

Making the pieces fit

Let’s stick with the party: this is an invented example of dividing up work, not a test of these products. One assistant looks for venues, another proposes a schedule and a third prepares the invitation. That could give you several initial options to make planning easier.

They all need the same basic details: date, guest count, budget and duration. Otherwise, one assistant might suggest a venue for two hours while the schedule needs three. Or the invitation might give a date that has not been confirmed yet.

A shared list helps: what is confirmed, what is a suggestion and what will you decide later? That makes the individual results easier to bring together and check.

Begin with a manageable task

Your first attempt could be: “Compare three venues for a party with 20 guests. We need a venue on this date for three hours and have this budget. Include sources and open questions. Do not contact anyone or book anything.”

Replace the date and budget with your own details. The task has a clear endpoint: you receive a comparison and can decide which venue to look into. Preparing a schedule or an invitation afterwards remains a separate step.

The same principle helps with research, writing or a software task: describe the result you need, the requirements and the decisions that stay with you. A well-defined task is useful even when only one assistant is working on it.

How to tell whether it helps

Tasks you know well are the best place to try this. You can judge whether a suggestion is useful and spot missing details. Perhaps you regularly compare offers or organise material for an article.

Afterwards, look at the whole result. Do the pieces fit? Can you trace the sources and understand the suggestions? How much time do you spend correcting things? That helps you work out whether the support actually saves you work.

If bringing the pieces together takes more work than dividing the task saved, the approach was not helpful for that task. More assistants working at once do not, by themselves, mean a better result.