---
title: "Personal AI agents compared: four runtime teardowns"
description: "Personal AI agents compared: Pine, Instinct, Muse and Town taken apart on the same six questions, from where the model runs to what memory keeps."
url: https://personalagentbench.com/agents/
last_updated: 2026-10-16
---

# Personal AI agents compared: four runtime teardowns

Personal AI agents compared here are products that keep state about you and act on your accounts: Pine, Instinct, Muse and Town. Most AI assistant comparison pages line chatbots up by features and price and then pick a winner. This one compares the four by how each one is built, and it doesn't pick a winner. Every product describes itself, so these pages do it the other way round. Each one is opened up at the runtime level, and every page answers the same six questions in the same order, which is the only way separate write-ups become columns you can actually read across.

- One page per product: Pine, Instinct, Muse and Town. Same six axes on each, same order, so the pages are comparable rather than merely similar.
- Everything on them was found by using the products patiently. No special access, no insider, nothing that required a login we weren't entitled to.
- Each page ends with what we failed to establish about that product, and those sections are not short.
- No page ranks its product against the other three. Completion rates live on the scorecard with their intervals attached. They were measured on our own replicas of each runtime with one substituted exam model, not on the consumer apps, and the teardowns stay a separate kind of evidence.

## The four personal AI agents on this bench

- [**Pine**Phones companies on your behalf about bills, refunds and cancellations.](https://personalagentbench.com/agents/pine/)
- [**Instinct**Lives in your messaging apps rather than in a new interface.](https://personalagentbench.com/agents/instinct/)
- [**Muse**Meta's personal agent, with a Linux machine of your own behind it.](https://personalagentbench.com/agents/muse/)
- [**Town**Web and native Mac, with the broadest native reach into your Mac of the four.](https://personalagentbench.com/agents/town/)

| Product | Examined surface and access | Price snapshot source | Principal limit |
| --- | --- | --- | --- |
| Pine | Web/Mac and business calling | [Dated plans and fees](https://personalagentbench.com/agents/pine/#pricing) | Server-side fact deletion unverified |
| Instinct | Messaging; invite or waitlist access | [Free for invited users; dated source](https://personalagentbench.com/agents/instinct/#pricing) | Main runtime host and model ID unknown |
| Muse | Web; per-user Linux VM | [Free use and paid-plan sources](https://personalagentbench.com/agents/muse/#pricing) | Personal memory-file access and Mac route untested |
| Town | Web and native Mac | [Free and paid tiers](https://personalagentbench.com/agents/town/#pricing) | Reach depends on permissions and configuration |

They were picked because they're all trying to do the same job and have made genuinely different bets about how. Pine phones companies for you. Instinct moved into WhatsApp instead of building an app. Muse hands every user a Linux machine. Town put a native process on your Mac with reach into your messages and your screen.

That spread is the point. A comparison of four products that all made the same architectural choice would tell you nothing, and most feature tables end up being exactly that.

Plenty of things get called personal agents and aren't on this list. Web search alone does not establish persistent state or connected-account actions; those capabilities depend on the product and mode examined. We wanted products that hold state about you and take actions on your behalf, since that combination is what creates every interesting question on the scorecard. The four here clear that bar.

One exclusion is ours rather than the category's. Zentor builds agent software, so Zentor's products are not on this bench and won't be added. Scoring yourself modestly and hoping the reader believes you is not a solution to a conflict of interest, and the [about page](https://personalagentbench.com/about/) spells out the rest of it.

Coverage reflects the products we could examine, not a ranking of the whole category. The [about page](https://personalagentbench.com/about/) explains the evidence boundaries and the differences between a teardown, a public-record page and an exam result.

## Which personal AI agent fits which job

There is no winner here, but there is a fit. Matching the job to the architecture is the short version of every page below.

- **You want a company phoned for you about a bill, a refund or a cancellation.** That is Pine's job. Calls placed from the app keep a summary rather than a verbatim transcript, which matters if you need to prove later what was agreed. Pine's own pages put personal plans on a weekly price with a monthly credit allowance; the dated figures are on [its pricing table](https://personalagentbench.com/agents/pine/#pricing).
- **You live in WhatsApp, SMS or iMessage and do not want another app.** That is Instinct's bet. A clock or an event on your mail, calendar or documents can start it, and secrets sit in a vault it can fill into a form but not read back. In [The Verge's hands-on](https://www.theverge.com/tech/1008254/instinct-agent-ai-hands-on-muse-dots), access meant an invitation from a current user or a place off the waitlist.
- **You want an agent with a machine of its own.** Muse gives each user a Linux VM and keeps your profile as markdown files on it. [Meta's launch post](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/) says it is free for most of what people need, with subscription plans for more. Its Mac app, which acts in your own apps, is the part we have not examined.
- **You want an agent on your Mac that reaches your messages, contacts and screen.** That is Town, the only one of the four where we confirmed a native foothold in the system. Routines can start on a schedule, an email or a calendar entry. It has a free tier.

If none of those is your job, [the landscape list](https://personalagentbench.com/landscape/) records the products we know of, and each task page, such as [disputing a charge](https://personalagentbench.com/tasks/dispute-a-charge/), says what to check before you hand an errand to any of them.

## What a personal AI agent teardown covers, and what it refuses to

Each page walks the same six axes: where the model runs, how a tool definition reaches it, where your memory is kept, what the product asks you to approve before it acts, which surfaces you can reach the thing on, and what we couldn't work out.

How the pages were made is its own page, [how we take them apart](https://personalagentbench.com/how-we-take-them-apart/), and the short version is that it involves getting a real account (paid where the product charges) and using the product the way a customer would rather than anything clever. The refusals matter as much as the coverage. We publish findings restated in our own words and never the material behind them, so you won't find instruction text, internal documents or schemas anywhere on this site, and you won't find a description of how anything was obtained either. If a fact can only be conveyed by reproducing somebody's prompt, it doesn't go on the page.

That rule costs us a few vivid details. It also means every page here can stay up.

One more thing is deliberately absent. We don't rank these products against each other, on any page, including this one, and each teardown shows its product's result status, with an interval and task count where a rate is available. Results and teardowns answer different questions, so the scorecard keeps them side by side rather than folding one into the other.

## ChatGPT, Claude and Gemini: where chat assistants fit in an AI assistant comparison

Most people searching for an AI assistant comparison want ChatGPT, Claude, Gemini and Copilot lined up as chatbots. We don't do that, for the reason above: an answer that ends with the conversation is a different job from an agent acting on your accounts.

That line has started to move. Several of those companies now sell an errand-running product beside the chatbot, and those are on [our list](https://personalagentbench.com/landscape/):

- OpenAI's [Dots](https://personalagentbench.com/openai-dots/), which Every's [side-by-side comparison of personal agents](https://every.to/personal-agents-comparison) describes as an always-on agent in ChatGPT.
- Microsoft's [Copilot Tasks](https://personalagentbench.com/copilot-tasks/). Microsoft's [help page](https://support.microsoft.com/en-us/microsoft-copilot/using-copilot-tasks) says it asks for approval or hands control back before sensitive actions such as payments, and that personal-account users will be moved to Copilot Cowork.
- Anthropic's [Claude Dispatch](https://personalagentbench.com/claude-dispatch/). Anthropic's [help page](https://support.claude.com/en/articles/13947068-assign-tasks-to-claude-from-anywhere-in-cowork) says you message Claude from your phone and it works on your desktop computer, and that Dispatch isn't available to new users.
- Google's Gemini Spark, which the same Every table describes as ongoing tasks across Google apps.

None of these has been through our teardowns or the exam. Their pages record what the vendor and the press have published, with each source named. If one gets a column here, it will answer the same six questions as the four above.

## Read them in whatever order, but start with the runtime

If you only read one section of one page, make it the execution section. Where the model runs decides almost everything downstream of it: what the agent can reach on your hardware, what a hostile web page could talk it into doing, and whether any record of what happened survives.

None of the four runs a model on your machine, which sounds like a tie and isn't. Pine assembles the whole request server-side and hands the client results, so the instructions shaping its behaviour never reach you. Instinct orchestrates from a server and does the work in a sandbox. Muse gives each user a Linux VM and builds the request on it. Town syncs through a backend while a native Mac app sits on your laptop with real system access.

That last one deserves a second read. The model isn't on your laptop, but a process that can see your messages, contacts, screen and audio is, and it takes its instructions from the model that isn't.

## An AI assistant comparison by architecture: where the four disagree most

Tools, memory and the trigger surface are where these products stop resembling each other.

On tools, you get four incompatible answers to one design question. Pine lets 26 tool names reach the client and no schemas at all, so you learn that a capability exists and nothing about its edges. Instinct ships a command line rather than a function list and won't run an action until its help text has been read, which puts the manual inside the prompt and inside your bill. Muse splits permission from definition: a sheet of 246 methods carrying no parameters, with the real definitions loaded a namespace at a time. Town holds a catalogue of 1105 and narrows it in four stages, showing its main assistant 177.

On memory, the practical question is what happens when the thing believes something wrong about you. Muse keeps your profile and notes as plain markdown in a home directory. Town keeps a wiki, a memory store, a library of documents and a people model server-side, with the personality as a document you can open. Instinct runs three stores for three jobs, one of which fills a form with a credential without revealing it. Pine keeps a structured fact layer you can query through the product and cannot open.

On triggers, at least two of the four, Town and Instinct, can start work without you typing anything. A clock, an inbound email, a calendar entry, a document change. An agent that can be started by an inbox is an agent a stranger can address, which makes the check between "something happened" and "act on it" more consequential than any feature on the table.

### The tool count trap

Every page here states a tool count, and every one of those numbers is close to useless on its own. We publish them anyway, with the caveat attached each time.

A catalogue number counts what exists somewhere in the product. What you want to know is how many tools the assistant in front of you can actually reach after you've connected only the accounts you were willing to connect. Those are different numbers, sometimes by an order of magnitude, and no product on this bench publishes the second one.

Town is the clearest case because it shows its working: 1105 in the catalogue, 177 offered to the main assistant, 17 for one background writer. Instinct is the murkiest, with three different totals depending on which part of the product is counting. Treat all of them as a rough sense of surface area rather than a figure to compare.

## Every page ends with a list of failures

Each teardown includes a section and an axis stating what we couldn't establish, and it's the section we'd point a sceptical reader at first.

We never saw Pine's production instruction text across roughly 126 probes, and we don't know which carrier places its calls. We got the section headings of Instinct's main instruction set and never the body, plus four named policy sections we found pointers to and no text for. With Muse we know the structure of what goes to the model and never saw one request leave the machine. Town's main template came back with 11 sections, and a fuller version said to run to 24 never did.

None of that is in a footnote. It's a row on the scorecard in the same typeface as the rows that flattered us, because a gap you have to hunt for is a gap somebody is hoping you won't find.

## How current these personal AI agent teardowns are

Every teardown carries the date it was last verified, and the scorecard carries one date for all four.

Products change. Pricing changes fastest, which is why current figures live in a dated data file on each product page rather than in prose where they'd quietly rot. Behaviour changes more slowly but it does change, and a teardown written against a version that no longer ships is a confident description of nothing.

If something here has gone stale or was wrong to begin with, the [corrections page](https://personalagentbench.com/corrections/) lists every fix with its date, and [about](https://personalagentbench.com/about/) says where to send one.

## FAQ

**What is a personal AI agent?** Software that acts across your accounts and devices rather than only answering questions: booking, cancelling, disputing a bill, keeping a calendar, running something on a schedule. The four on this bench are the ones we have taken apart; [the landscape list](https://personalagentbench.com/landscape/) records the rest.

**Which of these four personal AI agents is best?** We don't answer that, and the omission is deliberate. There are completion rates on the scorecard, from a September and an October batch, measured on our replicas of each product's runtime rather than on the apps. Some cells are empty on purpose: Town's September figure is withheld because the number we were given disagrees with the records we can check, and Pine has no October figure because almost none of its submissions reached the exam. No rate tells you what a product can reach on your machine or whether you can correct what it believes about you, which is what these pages are for.

**How were these teardowns done?** By using the products carefully and writing down what they did. No privileged access and no insider. The surface is the product, and a patient user can see a surprising amount of the machinery from outside it.

**Why don't you publish the material behind the findings?** Because publishing somebody's instruction text is republishing their work, and because a site that did it would be arguing about takedowns instead of doing teardowns. Findings in our own words, never the artefacts.

**Do any of these agents run the model on my own device?** None of them. Muse gives you a Linux VM of your own, Town puts a native Mac process with system access on your laptop, Pine hands browser work to a local agent on the Mac app, and Instinct runs delegated work in a cloud sandbox; where its main agent runs we could not establish. The model itself is remote in all four cases.

**How often are the teardowns updated?** Each page carries its own last-verified date. We re-check when a product ships something that changes an answer on one of the six axes, and we'd rather show you a stale date than let a page imply it's current when it isn't.

## Comparing personal AI agents on how they work

Each of these four is good at something and opaque about something else, and the pattern of which is which is the interesting part. Pine does one narrow job thoroughly and shows you almost nothing about how. Instinct moved out of the app and into your message threads, and kept a careful line between using a credential and seeing it. Muse keeps a parameter-free permission sheet apart from its tool definitions, and the traffic between its VM and the model is the one thing we never saw. Town's Mac app reaches further into the system than the other desktop apps we examined, and it carries a lot of machinery for limiting what it does there.

The [recorded runs](https://personalagentbench.com/runs/) publish individual exam attempts a page at a time, with the conversation, what changed in the simulated world, and the measurements that came out of it. Those attempts come from our own replicas of each product's runtime, run on one substituted exam model, so they show how each arrangement copes with an errand, not a recording of the app you would download.

These four are not the whole category, and [the landscape list](https://personalagentbench.com/landscape/) says so: it records every personal agent we know of, with the four here marked as the ones we have taken apart and whose runtime replicas are in the exam, and most of the rest marked as recorded and nothing more.

[The comparison pages](https://personalagentbench.com/compare/) sit alongside these, one for every pair of products we have opened, each putting two runtimes side by side on the same axes. We only build them for pairs we have opened ourselves, because a comparison with marketing on one side is the thing this site exists to replace. [OpenMuse](https://personalagentbench.com/openmuse/), an open source personal agent unaffiliated with Meta, is the control case: we read its source rather than probing it. [Where it runs](https://personalagentbench.com/where-it-runs/) takes the first axis across everything we know of, [who acts unasked](https://personalagentbench.com/proactive-ai-agents/) asks which of the four start work before you do and whether they check first, and [Grok Bot](https://personalagentbench.com/grok-bot/) records what is public about a product that is in the exam but has not been taken apart.

Pick whichever product you're closest to choosing and read its page top to bottom, then read the same section of one other. After that, [the scorecard](https://personalagentbench.com/) puts all four side by side, [how we score](https://personalagentbench.com/how-scoring-works/) explains where the axes came from, and [the benchmarks section](https://personalagentbench.com/benchmarks/) covers why the existing agent leaderboards don't answer this question.

---

Published by Personal Agent Bench (https://personalagentbench.com), run by the Zentor team. Zentor builds agent software of its own, so its products are not on the scorecard and will not be added.