Muse works while you're away

A chatbot only runs while you're typing to it. Real errands don't work that way. Haggling over a bill can take days of back-and-forth, and a good flight price might not show up until next Tuesday. Meta says that for longer tasks, "Muse keeps working after people close the app, and comes back when something changes or when it needs approval."

For that to be possible, the agent needs somewhere to live that isn't your phone. Scroll to see where.

A computer of its own

Meta's design gives every user a dedicated virtual machine in the cloud. In their words, "You and your Muse share your own dedicated computer." It has real storage, CPU and memory, enough to run code the agent writes for itself, keep background jobs going and run several helper agents at once.

That's a big change from a chatbot, which is closer to a very clever function: text in, text out, then nothing. An agent with its own computer is more like a process that stays alive. It can wait, get woken up, check something and go back to waiting.

Scheduled jobs: waking itself up

Some tasks are about noticing when something changes: "tell me when this drops under $200", "remind me before Sam's birthday", "check whether the refund arrived." Muse's VM supports crons, the old Unix name for scheduled jobs. A cron wakes the agent at set times, and it runs the loop from Chapter 1, looks around and then either goes back to sleep or pings you.

Subagents: many hands at once

Big jobs split naturally into smaller ones. Planning a weekend away means checking flights, hotels and restaurants, and none of those depends on the others. So the main agent acts as an orchestrator. It breaks the goal into subtasks and hands each one to a subagent, a fresh copy of the agent with its own narrow job and a clean context. Meta says Muse can "launch swarms of subagents."

This helps in two ways:

• Speed. Three searches running in parallel finish in about the time of one, the same trade-off as Chapter 2's Contemplating mode, applied to real work.
• Focus. Each subagent sees only its own slice of the problem, not a huge transcript full of unrelated flight data. Smaller, cleaner contexts lead to better decisions and cheaper calls.

Meta also describes what makes a good subagent: it "adheres to its job, understands available tools, and knows when to escalate back to the main agent." A subagent that wanders off-task or quietly makes a big decision itself is a liability.

Interrupt only for decisions

A background agent is only pleasant if it's quiet. It shouldn't send updates every few minutes or ask whether you're sure about Chicago. The design goal is to come back only when something changed (the fare dropped) or when a decision needs you (spending $682).

That second kind of interruption, the approval, turns out to be one of the most important security features in the whole system. Part 3 explains why.