Personal AI assistants, benchmarked and taken apart
We give the personal AI assistants people hand real errands to an exam, and we take each one apart to see how it works underneath. Rates come with their 95% interval and sample size. There are no ranks and no combined score.
Share of errands completed
11 October: all four given the same 20 tasks, once each. Not comparable with September.
Muse58% (36–77)
Instinct40% (20–64)
Town15% (5–36)
PineNo data, exam fault
The 09-21 to 09-22 records, filtered, 2026-09-22. All four answered the same tasks.
- Instinct52% (45–61)
- Muse40% (33–46)
- Town37% (30–44)
- Pine30% (23–36)
As of 2026-09-22. Each product answered its own draw, so rates are not directly comparable.
- Instinct43% (36–50)
- Pine20% (14–25)
- Muse19% (13–25)
- TownFigure withheld
GrokExam pending
95% interval measured rateWhere two bands overlap, the gap is within measurement noise. These are product runtimes rebuilt for the exam with one substituted model, not the consumer apps measured end to end. Full tables, caveats and data
- 4runtimes taken apart
- 6axes, same for each
- 416scored product-task observations
- 2026-10-09scorecard checked
Four personal AI assistants, taken apart
A completion rate says how often something finished the job. A teardown says what it can reach, what it keeps about you, and what you can check afterwards. Each of the four has its own page answering the same six questions in the same order.
| Personal AI assistant | What it does | Made by | Price |
|---|---|---|---|
| Pine | Phones companies for you about bills, refunds and cancellations | 19Pine (App Store) | Published: weekly personal plans, monthly work plans |
| Instinct | Lives in WhatsApp, SMS and iMessage | Spear Street Technology; invite-only since August 2026 (TechCrunch) | No published price |
| Muse | Meta's agent, with a Linux machine of your own behind it | Meta (launch post) | Free for most use, with subscription plans |
| Town | Web and a native Mac app with wide reach into your Mac | Town.com | Published, starting with a free plan |
Current prices, dated, are on each product's page.
Six questions for every personal agent
Each row is a question worth asking before you hand a product your accounts. The short answers are below; the full ones are on each product's page, and how we score explains where the six came from. The last row, what we could not establish, is the one we would point a sceptic at first.
| Axis | Pine | Instinct | Muse | Town |
|---|---|---|---|---|
| Where it runs | Server-side; browser work runs on your Mac | Server plus sandbox; leased cloud browser | One Linux VM per user | Server-side; native Mac reaches the system |
| How tools are exposed | 26names reach you, no schemas | 571by its own count; manual read enforced | 246permissions; schemas loaded per group | 1105listed; 177 offered to the assistant |
| Where memory lives | Server-side facts, no file to edit | Git-tracked markdown, indexed connector data, fill-only vault | Plain markdown in a home directory | Server-side wiki, memory, library and people |
| What you get to approve | Two confirmations: start, and identity | Approve, deny or dismiss; bound to exact text | One gatekeeper for approvals and egress | Approval set by configuration, not a prompt |
| How you reach it | Web, Mac, outbound calls, REST API | WhatsApp, SMS, iMessage, its own email | Apps, web, WhatsApp; Mac untested | Web and Mac; clock and email triggers |
| What we could not establish | Instruction text, call schema, carrier | Instruction body, host, model id | Everything sent to the model | Instruction set, personality placement |
Where personal agents differ most
- Where it runs. None of the four runs a model on your machine, but Town puts a native Mac process with access to your messages and screen on your laptop, and Muse gives every user a Linux VM. Where it runs takes this across every agent we know of.
- Where your memory lives. Muse keeps your profile and notes as plain markdown; Pine keeps a structured fact layer with nothing you can open. If it believes something wrong about you, the fix ranges from "edit a file" to "go without".
- What can start it. Town and Instinct can start work without you, from a clock, an email or a calendar entry. An agent an inbox can trigger is an agent a stranger can address.
Beyond the four personal AI assistants
- All personal agents we track, including Dots, Copilot Tasks, Claude Dispatch, Poke and Grok Bot, each with its evidence.
- Open-source personal agents, read from source: OpenClaw, Hermes, OpenMuse and more.
- Phone use agents that operate a phone for you, compared on how they reach the screen.
- Head-to-head comparisons for every pair we have opened, and recorded runs of single attempts.
- What other agent benchmarks measure, and why none answers this question.
What this personal agent benchmark does not do
It does not rank products or add the rows into one score, because that means deciding for you how much each one matters. It does not score price or output quality; how we score says why. And it never includes Zentor's own products: Zentor, the team behind this site, builds agent software, so it stays off the bench. The about page has the rest.
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 on this scorecard are the ones we have taken apart; the rest are under all personal agents.
Which personal AI assistant is best? We do not answer that, deliberately. Read the rates with their intervals, then read what each product can reach and keep, and weigh them yourself. In the first batch, one product's figure is withheld because the number we were given does not match the records we can check; the results page explains it.
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 runs delegated work in a cloud sandbox; where its main agent runs we could not establish.
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.