Why Agents Became Personal (And What Gets Built Because of It)

Agents became personal when three things got cheap enough to hand to every user at once: a computer for the agent to work on, a safe place to keep that user's credentials, and a way to pay. Muse ships all three by default, a dedicated VM, a surrogate-token vault and a Stripe Link wallet, and that is why the builds in this catalog look like errands and plumbing rather than chat tricks. The model matters, but it was not the thing holding agents back.

This is an argument, not a news post. Every fact in it links to a source, and every example is a real entry in the catalog.

The scarcity we stopped noticing

For two years the conversation about agents was about intelligence. Could the model plan? Could it recover from a failed tool call? Benchmarks kept score, and they still do: Meta reports Muse Spark 1.3 at 75.4 on DeepSWE v1.1 and 66.9 on OSWorld 2.0 at max effort (MarkTechPost).

But intelligence was never the scarce input for a personal agent. The scarce inputs were boring. A machine that stays on while you sleep. Logins to your bank, your email and your kid's school portal. A card number you are willing to let software use. Developers had all three for themselves, which is why the first good agents were coding agents. Ordinary people had none of them in a form an agent could safely touch.

Muse's launch post describes an agent that "keeps working in the background after the app is closed" (Meta). That sentence is about infrastructure, not intelligence.

A computer per person, not a chat window

Each Muse user gets a dedicated Muse Secure VM built on systemd-nspawn runtime cells (Meta research). Greg Ceccarelli got Claude Code running inside his and reports about 2 vCPUs, 8GB of RAM and 100GB of disk, with no inbound ports. Peter James had the iOS app zip the whole filesystem and found a home directory with soul.md, identity.md and memory files.

A computer changes what people ask for. You do not ask a chat window to watch Fandango prices for a week. You do ask a machine that never logs off. One user had Muse price books from photos, list them on eBay, and keep tracking a water-filter schedule. Another had it play Sonic on an emulator it downloaded itself. Neither request makes sense without a persistent machine.

The builder response was to connect that machine to the rest of people's lives. One Redditor joined the Muse VM to his Tailscale network so it could SSH into a restricted user on a home Ubuntu box. TerMuse exists so you can watch and take over the agent's terminal and browser side by side.

Credentials were the real bottleneck

An agent that can reach your accounts can also leak them. Muse's answer is credential surrogation: the agent's code only ever sees a stand-in token, and a separate authority called Sentinel swaps in the real secret at the network edge and gates connector actions and egress (Meta research).

That one design choice is what makes custom connectors safe enough to ship to consumers. Muse can write an integration for any service with an API, and the key goes into a secure store instead of the chat (Meta help). A Redditor built a Linear connector in under a minute. Builders then went further and made the vault the product: Muse Proxy stores CalDAV passwords once and issues scoped keys, and the Cloudflare Workers starter keeps upstream API keys server-side behind one bearer token.

The pattern is consistent across the connector builds. People are not building clever prompts. They are building narrow, auditable doors.

Payments made errands finishable

An errand is not done until something is paid, booked or cancelled. Muse launched with Stripe Link, which checks out with a saved method at more than 1M Link merchants and issues a single-use virtual card everywhere else (Stripe). Within two weeks Shop Pay, PayPal, Expedia and Instacart were announced as partners.

That is the difference between an agent that finds things and one that finishes them. Alex Volkov gave Muse and another agent the same job, tickets for a Rosh Hashanah dinner, and Muse booked through Link while the other was still searching. Ejaaz asked for compensation on a 7-hour Delta delay and, by his account, had a $250 credit five minutes later.

Meta has said how it plans to make money from this: "by taking a small fee from transactions" (TechCrunch). The business model and the product point the same way, toward the agent completing purchases.

What gets built when agents are personal

Look at what the first weeks of builds actually do. A lot of them find money. Dowser surfaces class-action settlements and expiring points. A receipt audit found a forgotten $100/month subscription. An unclaimed-property search turned up $88.92. The team behind Agents Breakroom sorted about 500 real Muse use cases and found 20% were about money.

A lot of them talk to the phone system. Two separate builders shipped phone numbers for Muse, one for inbound and outbound calls and one where Muse whispers reminders mid-call. Matt Schlicht's Muse called five Tesla stores at once.

And a lot of them are about the household. School and sports feeds, a family morning newsletter in How I AI's review, 18 Spotify playlists re-sorted by decade. None of this is technically impressive. It is personal, and it only works because the agent holds the logins. See real errands people delegated for the full list.

Where the personal model breaks

Personal agents inherit personal problems. Amazon blocked Muse purchases from September 20, saying the agent accessed its store without permission and did not identify itself (GeekWire). Within two days a builder had a self-hosted cart API Muse reaches over SSH, which tells you how much people want the errand done.

Patrick Wardle reported a Mac zero-day where a local app can redirect dictation traffic and expose auth tokens (Malwarebytes). Taylor Arndt found VoiceOver gaps that kept her from using it for work. Muse is US-only and 18+, and TechCrunch asked the obvious question at launch: will consumers trust Meta with this (TechCrunch)?

Joshua Liu said Muse saved him $113 a month on car insurance and in the same post worried agents could widen the digital divide. Both halves are true. An agent with your logins and your card saves money for the people who have one.

The takeaway for builders

If you are building for Muse, build for the scarce inputs. Connectors that keep keys out of the model. Bridges from the cloud VM to hardware people own. Read-only tools that surface money or deadlines. The connector platform had more than 1,500 applications in under a week (TechCrunch), so the easy integrations will be crowded. The personal, specific ones are not.

Frequently asked questions

What makes an AI agent "personal"?

A personal agent acts on your own accounts, with your own credentials, on a machine that keeps running when you are not watching. Muse provides this with a dedicated VM per user, a surrogate-token credential store and a payment wallet through Stripe Link.

Why did coding agents come before personal agents?

Developers already had the ingredients personal agents need: an always-on machine, API keys and tolerance for risk. Consumers needed those supplied safely, which is what Muse's VM, Sentinel authority and Link integration try to do.

Is Muse safe to give my passwords to?

Muse is designed so the model never sees your real secrets; Sentinel swaps them in at the network boundary. It has still had problems, including a reported Mac zero-day and Amazon's complaint about how it accessed its store, so start with read-only connectors and approvals turned on.

What do people actually use personal agents for?

In this catalog, mostly money and admin: bill negotiations, refunds, subscription audits, unclaimed property, bookings, and family schedules. Many of the agent builds report recovering or saving a specific dollar amount.