Three Background Agents Now Compete to Initiate Your Work: Run the Delegation Audit First

The category is no longer optional

Three of the biggest AI vendors now ship roughly the same thing: a background agent that initiates work instead of waiting to be asked.

Meta announced Muse first, describing it as a personal agent that carries out long-running tasks and keeps working in the background. OpenAI's Dots can text you proactively when something you asked it to watch — a product's price, a travel itinerary, a deadline — changes. And on October 21, Anthropic opened a research preview of Claude Duet: a proactive agent built around Opus 5.5, with access to the over 6,000 connectors in the Claude Marketplace, a morning brief of what needs attention, and a rule that it checks with you before sending a message, making a payment, signing anything, or deleting anything.

The competitive coverage treats this as a horse race — who answered whom, which vendor is ahead. That's the wrong read for a business owner. The real change: agents that come to you are now table stakes, so the interesting decision is no longer whether to use one. It's which of your recurring tasks should hand off — and that's a decision to make before you enable anything, not after.

The chat-agent question was "can I trust it with a task I assign?" The background-agent question is different and harder: "which of my tasks should an always-on agent own?" Here's the audit to run first.

Step 1 — List what hands off cleanly: trigger-checkable work

Not every recurring task is delegable, even to a competent agent. The test is the trigger. A task hands off cleanly when all four are true:

1. It recurs on a pattern, not a deadline. Invoice chasing, review responses, appointment follow-ups, price-drop watches. Not "redo the website."
2. The trigger lives in a connected system. The agent's superpower is checking accounts on a schedule — Duet checks connected accounts every few hours, Dots watches what you hand it. If the trigger is visible in email, a bookkeeping app, a CRM, or a review site, the agent can see it start. If the trigger lives in a phone call or a hallway conversation, it can't.
3. A wrong answer is cheap to catch. A draft reminder you read before it goes costs nothing. A misfiled tax election doesn't.
4. Nobody at the business enjoys doing it. Low morale, high frequency — that's the sweet spot.

Anthropic's own task list for Duet is a useful map of what passes this test: finding missed Gmail and Slack messages and helping reply, checking QuickBooks or Stripe for unpaid invoices and drafting reminders, drafting follow-ups from HubSpot deal history. Notice what those have in common — read, draft, prepare, then stop for approval. That's the shape of a first safe delegation.

Step 2 — Write the never-hand-off line before the agent offers to cross it

Background agents can do more than read-and-draft now — Duet can prepare payroll in Gusto and, with your approval, pay with a virtual card through Link. That's the point of the product. It's also why the boundary has to be written down first, while nothing is connected.

The line worth keeping: money moving and promises made stay with a human. An agent can draft the payment; a person releases it. An agent can draft the client reply; a person sends it. Duet's own design concedes this — before it sends, pays, signs, or deletes, it shows you what it's about to do and waits. A few things are off-limits even with approval, like trading securities or moving money between accounts.

Vendors building the boundary in is good. It's not the same as you deciding where your boundary sits. Write the list before the first connection, because the moment an agent surfaces a half-finished tempting action is the wrong moment to invent policy.

Step 3 — Set an interruption budget before week two

Here's the failure mode nobody covers in the launch posts: the agent that interrupts more than it saves.

A background agent's whole mechanism is interrupting you — that's what "proactive" means. A Dots agent texting about a price drop you no longer care about, a Duet ping about an email you'd have ignored anyway — each one is individually reasonable and collectively a second inbox. Anthropic's announcement is unusually candid about the tradeoff: Duet checks your connected accounts every few hours because checking nonstop would use up your plan's limits much faster, and testers are told "it still might misread a thread or miscalibrate what's urgent."

So treat notifications like a budget. Rule of thumb: if a delegated task generates more than one interruption a day, the task wasn't delegable — or the notification threshold is wrong. Fix the threshold once, and if it pings twice, revoke the task. The task list should shrink toward the few that pay for themselves, not accumulate. A weekly review of what the agent actually surfaced — not what it says it did — is the companion discipline once things are running.

Step 4 — One owner per surface, or three agents will fight for your background

If you trial all three (don't — see boundaries below), the practical rule: never point two agents at the same app. Muse watching your email, Dots watching your email, and Duet drafting from your email is three sets of credentials, three notification streams, and a real chance two of them act on the same thread. One app, one agent, one owner — noted in the same place you wrote the never-hand-off list.

This sounds obvious and it's the first thing that breaks, because each agent's setup flow encourages connecting your whole stack on day one. Connect only the app that feeds the one task you're trialing.

The 30-minute version, tonight

1. List five recurring tasks with triggers in connected systems (Step 1's four tests — most owners find two of five pass).
2. Write the two lines: the never-hand-off line, and the one-owner-per-app rule.
3. Pick one task on one agent, on a two-week trial with an interruption budget of one ping a day.
4. Diary the revoke date. If at two weeks the interruptions outweigh the catches, disconnect the task — that's a successful audit, not a failed experiment.

Honest boundaries

• Everything above is a research preview or a first release. Duet is waitlist-only, US, on paid Claude plans, and shares usage limits with the rest of the product — background checks count against your plan. Muse and Dots are consumer-first personal assistants; Duet is deliberately work-focused. Expect churn in all three.
• These are subscription products with usage costs, not one-time purchases. A background agent that checks four apps every few hours, every day, is a line item that grows. Price the trial accordingly.
• Don't rebuild your business around the tool. Vendors will iterate, merge, and rename these products inside a year. The audit — what recurs, what never hands off, what the interruption budget is — transfers to whichever agent wins. The tool choices don't.
• The audit is pre-delegation. It decides what starts. It's the pair of, not a replacement for, a weekly review of what's already running.

---

Want the delegation worksheets? The Boring Automation Pack includes the recurring-task inventory and the permission-line worksheet — the boring half of this audit, pre-built.