Four personal agents, taken apart at the runtime level

Every personal AI assistant comparison you can find counts features. A feature list tells you what a product claims. It tells you nothing about what happens after you press send, which is where these four products differ most and where none of them competes on the marketing page.

• 4runtimes taken apart
• 5axes, same for each
• 0scores published
• 2026-09-20last verified

What this scorecard says

Four products, examined at the runtime level rather than by clicking around. Five axes, the same five for each. No scores, because scores have to come from recorded task runs and those runs have not happened yet.

The last row is the one nobody else publishes: every column states what we failed to find out. And everything here is a finding restated in our own words, never the material behind it, so you will not find instruction text, internal documents or schemas on this site.

How to read it

Each row is a question you would actually ask before handing a product your accounts, and each cell is a short answer with a full one on the product's own page.

The rows are deliberately not weighted against each other. There is no combined score at the end, because combining "the model runs server-side" with "memory is a file you can open" into a single number requires deciding how much each one matters to you, and we do not know that. Anybody who does produce that number has decided it for you and not told you how.

Read across a row to compare. Read down a column to understand one product. Both readings are supported, which is the entire reason the table has this shape.

• PinePhones companies on your behalf about bills, refunds and cancellations.
• InstinctLives in your messaging apps rather than in a new interface.
• MuseMeta's personal agent, with a Linux machine of your own behind it.
• TownWeb and native Mac, with a real foothold on the machine itself.

Who this is for

If you are choosing a personal AI assistant for yourself, the useful question is not which one is cleverest. All four are built on frontier models and all four write a decent email. The question is what each one is allowed to touch, what it keeps, and what you can check afterwards, because those are the parts that differ and the parts you cannot change later.

If you are building one, this is a survey of four serious attempts at the same problem, with the design decisions made explicit. Some of them we would copy tomorrow. Some of them we would not.

And if you are simply curious how any of this works underneath the chat box, start with any teardown and read the runtime section. None of it requires special access. The surface is the product, and everything on this site was found by patient use.

Where the model runs, and why it decides everything else

None of the four runs a model on your machine. That sounds like a tie until you look at what each one does instead.

Pine builds the whole request server-side and hands your client the results, so the instructions shaping its behaviour never reach you. Instinct orchestrates from a server and runs the actual work in a sandbox. Muse gives every user a Linux VM and assembles the request there, on a machine that is in some meaningful sense yours. Town syncs through a shared backend while a native Mac app sits on your laptop with real system access.

That last arrangement is the one worth thinking hardest about. The model is not on your laptop. A process with access to your messages, contacts, screen and microphone is, and it takes instructions from the model that isn't.

Execution location decides what an agent can reach, what it can be made to do by a poisoned web page, and what evidence exists afterwards. Everything else on this table is downstream of it.

How tools reach the model

Four products, four completely different answers to the same design question, which is unusual enough to be worth stating plainly.

Pine exposes 26 tool names and not one schema, so you know a capability exists and nothing about its edges. Instinct exposes a command line and refuses to run an action until its manual has been read, which makes documentation part of the prompt and part of the bill. Muse splits the problem in half: a permission sheet of 246 methods carrying no parameters at all, and real definitions loaded a namespace at a time. Town lists 1105 tools and narrows to roughly 177 through four separate filters before the assistant sees any of them.

Only one of those four separates what an agent is allowed to do from how it does it. That separation is why the question "what can this thing touch" has a short answer on Muse and a long one everywhere else.

There is a trap in tool counts, and it applies to all four. A catalogue number describes what exists. What you want is what the assistant in front of you can reach after you have connected only the accounts you are willing to connect. No product on this bench publishes the second number.

Where your memory lives

This is the axis with the sharpest practical consequence, and the one most buried in marketing.

Muse keeps your profile and notes as plain markdown in a home directory, embedded straight into the request. Instinct keeps three separate stores doing three different jobs, including a credential vault built to fill a form without ever revealing what it filled. Town keeps a wiki, a memory store, a library and a people model, all server-side, with the personality as its own editable document. Pine keeps a structured fact layer queried on demand, with nothing you can open.

Ask one question of any personal AI assistant you are evaluating: if it believes something wrong about me, how do I fix it? On this bench the honest answers range from "edit a file" to "go through the product or go without".

How you reach it, and what happens on the surfaces you do not use

Pine ships web, Mac, outbound phone calls, a voice copilot and a REST API other people's agents can call. Instinct lives in WhatsApp, SMS and iMessage, with a web app that functions as an admin panel. Muse is web in practice, whatever the documentation says about other channels. Town treats web and a native Mac app as equals.

The interesting part is not the list. It is that three of the four can start work without you: a clock, an incoming email, a calendar entry, a change to a document, a location. An agent whose trigger surface includes an inbox is an agent that strangers can address, which is why the safety check between "something happened" and "act on it" matters more than any feature on this table.

What we could not establish

A bench that only reports wins is marketing, so the last row of the scorecard is the one we would point a sceptical reader at first.

We never saw Pine's production instruction text, across roughly 126 probes. We got the section headings of Instinct's main instruction set and never the body. We know the structure of what Muse sends to a model and never saw a single request leave the machine. We know Town's template runs to 24 sections and that the member-facing interface will not return them.

None of that is hidden in a footnote. It is a row in the table, sitting in the same typeface and the same column width as the rows that flattered us, because a gap you have to go looking for is a gap you are hoping nobody finds.

What this scorecard deliberately does not measure

Three things a reader might expect here are missing on purpose, and it is fairer to say which than to let you assume they were forgotten.

Price is not a row. All four products change their pricing faster than a comparison page can honestly track, and a stale price is worse than no price: it looks authoritative and it is wrong. Current figures live on each product page in a dated data file rather than in prose, so when they go stale they say so.

Output quality is not a row either. Whether an agent writes a good email or negotiates well is a judgement, it varies per task, and every attempt we have seen to score it ends up measuring the scorer. Runtime facts are checkable by anyone with an account and patience. We would rather publish the checkable thing.

Our own products are not on the table and never will be. Zentor builds agent software, which is a genuine conflict of interest, and the only honest way to handle it is to stay off the bench rather than to score ourselves modestly and hope you believe it.

Why there are no scores yet

Because we do not have any, and the alternative is making them up.

Scores have to come from recorded task runs: hand the same errand to all four, write down what happened, publish the runs that went nowhere alongside the ones that worked. Those runs have not happened yet. Until they do, this page carries a NOT YET SCORED mark and four columns of findings that each point at the evidence behind them.

A competing comparison currently shows 116 products against 15 dimensions. We cannot tell you where any of those cells came from, and neither can it. Four columns we can defend is the better trade.

FAQ

What is a personal AI assistant? Software that acts on your behalf across your accounts and devices rather than only answering questions: booking, cancelling, disputing a bill, keeping a calendar, running a routine on a schedule. The four products on this scorecard are the current serious attempts at it.

Which personal AI assistant is best? We do not answer that yet, deliberately. The scorecard has no scores because scores have to come from recorded task runs and those have not happened. What it does give you is four runtimes described on identical axes, so you can weigh the trade-offs that matter to you.

Do any of these agents run on my own computer? None of them runs the model locally. 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. Instinct keeps everything in a cloud sandbox.

Can I see what these agents were instructed to do? Mostly no, and that is what the last row of the scorecard records. Pine's instruction text never reaches the client at all. Instinct surfaces section headings only. Town's member interface refuses the full template.

Who runs Personal Agent Bench? The Zentor team, which builds agent software of its own. That is a real conflict of interest, so Zentor's products are not on the scorecard and will not be added. The disclosure sits in the footer of every page.

How often is this updated? Each product page carries the date it was last verified. Teardowns have a shelf life because products change, and we would rather mark ours than pretend otherwise.

Choosing a personal AI assistant by how it works, not by what it claims

Every product here is good at something and opaque about something else. Pine does one narrow job thoroughly and shows you almost nothing. Instinct made the boldest interface bet and the most careful security decisions. Muse has the architecture we would copy and the least visibility from outside. Town has the widest reach and the most machinery for limiting it.

Start with the teardowns, read how we score, or see what the existing agent benchmarks measure and why none of them answers this question.