About Personal Agent Bench

Personal Agent Bench is a public comparison of personal AI agents, four of them taken apart at the runtime level and more in the task exam, written by a team that builds agent software for a living. That's a conflict of interest and it's better said on the first line than discovered on the fourth. This site is personalagentbench.com. Exam results, teardown observations and vendor claims are different kinds of evidence.

• Built and run by the Zentor team. Zentor's own products are not on the scorecard and will not be added later.
• Every teardown finding was found by using the four products carefully; completion rates come from a separate exam. No privileged access, no insider, nothing behind a login we weren't entitled to.
• We publish conclusions restated in our own words and never the material behind them: no instruction text, no internal documents, no schemas.
• Completion rates are published, each with its interval and the number of tasks behind it. The September batch has no public record behind it; in the 11 October batch every attempt has a record, and we checked the figures against those records. There's still no combined score and no ranking, and a figure we can't reconcile with records we can check is withheld rather than printed with a hedge.

Who runs Personal Agent Bench, and the conflict that comes with it

The Zentor team. Zentor builds agent software, which puts us in the same category as the products on this bench and gives us an obvious reason to make some of them look worse than they are. That includes software that operates a phone: the tools on our phone use agents pages are direct competitors of work Zentor is doing, except one: Oh My Phone Use is a side project by people at Zentor. It is marked as such where it appears, read with the same checklist as the others and given no score.

These pages speak for Personal Agent Bench. The site does not publish named individual author biographies, so a reader cannot independently assess a particular writer's experience from a byline. Do not infer a credential, outside affiliation or endorsement from the organisational author field. The evidence on each page must carry the claim.

There is no clever way around that. The only handling that actually works is structural: we stay off the table. Zentor doesn't appear on the scorecard, won't be added in a later revision, and isn't quietly the implied winner of any comparison here. The disclosure sits in the footer of every page, not just this one, because a conflict declared once on an about page is a conflict most readers will never see.

Pages separate what we observed from what other people reported, with dates and open questions where we have them. Each results table says which batch a figure comes from and what record stands behind it. That helps you challenge the evidence. It does not remove the operator’s conflict.

If that's disqualifying for you, that's a reasonable position. Read the method and judge the work rather than the byline, or don't.

Why a bench for personal AI agents at all

Because the comparisons that existed counted features, and a feature list tells you what a product claims.

Try to answer an ordinary question from one of those tables. If this agent gets something wrong about me, how do I correct it? What can it reach on my laptop while I'm asleep? If a web page it reads tries to give it instructions, what stops that? None of those are features. They're consequences of how the thing is built, they differ enormously between products that look identical in a feature grid, and they're the parts you cannot change after you've committed.

Linking findings to evidence is worth doing. The September completion batch is the exception: no public batch record is available, as the scorecard explains. The 11 October batch is different, because every attempt behind it has a record we checked its figures against. Folding the results into one overall number is the part we won't do, because producing it means deciding for you how much each question matters.

What "runtime level" means in practice

It means we describe what happens after you press send, rather than what the marketing page says the product does.

Concretely: where the model executes and what sits between it and your data; how a tool definition gets into the context and who decided which ones; what the product stores about you, in what form, and whether you can open it; which surfaces the thing listens on and what can start it working when you aren't there; and what it stops to ask you about before doing something that cannot be taken back. Those five, plus an honest account of what we failed to find out, are the six axes on the scorecard.

They weren't chosen in advance. We took the products apart first and built the axes out of what turned out to be different, which is why the axes are oddly specific and why they're the same six for every product.

What Personal Agent Bench publishes, and what it doesn't

The line is simple to state and occasionally annoying to hold: findings yes, artefacts no. How we take them apart is the long version, including the three evidence labels we use internally and why two of them never reach a reader.

We publish what we concluded, in our own words. We don't reproduce instruction text, internal documents or schemas from any product on this bench, we don't describe how anything was obtained, and we don't publish a quote where a restatement will do. Numbers, structures and behaviours are fair game. Somebody else's sentences are not.

Two reasons, and the second is the load-bearing one. Publishing another company's prompt is republishing their work, whatever you think of the company. And a site that did it would spend its time arguing about takedown notices rather than doing teardowns, which would be a slow way to publish nothing.

That rule is enforced by a check that runs before anything goes live, not by good intentions on the day. Drafts are diffed against our own source material, and an overlap long enough to be a lift rather than a coincidence stops the publish outright.

Why there's still no combined score

Because a combined score needs weights we don't have, and the alternative is inventing them.

The first task runs exist now, so completion rates are published. What isn't published, and won't be, is one number per product. Folding a completion rate together with where the model runs, what it can reach on your machine and whether you can correct what it remembers requires deciding how much each of those matters to you, and that depends on your situation rather than on anything we measured.

The rates themselves are published the awkward way on purpose. Each one carries a 95% interval and the count of tasks behind it, and the table is ordered by sample size rather than by rate. A row can be greyed out for resting on too few tasks to mean anything, and a figure can be pulled from the table entirely when it doesn't reconcile with the records we're able to check, with the disagreement described in the cell where the number would have been. A single sorted column would be shared far more widely. It would also be the thing this bench exists to point at.

The last row is the product

Every column on the scorecard ends with what we couldn't establish, and those entries aren't short.

Pine's production instruction text never surfaced across roughly 126 probes. Instinct gave up the section headings of its main instruction set and never the body. With Muse we know the structure of what goes to the model and never saw a single request leave the machine. Town's main template came back with 11 sections, and the member-facing interface never returned the fuller version said to run to 24.

The same principle runs through the recorded runs, where an attempt that went nowhere gets a page of its own rather than being dropped from the sample. Those sit in the same typeface and the same column width as the findings that made us look thorough. A bench that only reports its wins is marketing with a table in it, and the failures are also the most useful signal on the page: they tell you exactly how much of each product is visible to a determined outsider, which is itself an answer to a question worth asking.

How Personal Agent Bench is paid for

It isn't, in any direct sense, and that's checkable rather than a claim you have to take on faith.

There's no advertising on this site, no affiliate links, no sponsored entries, and no tracking of any kind. Nothing here fires an analytics beacon, nothing sets a cookie, and no link out to a product carries a referral tag. Open the page source on any URL and you'll find static HTML with its styles inline and no third-party script.

The site is a cost that Zentor carries, which is a different conflict from the usual one and worth naming rather than presenting as purity. A comparison funded by affiliate commissions has an interest in you clicking through to whichever product pays best. This one has an interest in Zentor looking like the sort of company that publishes careful work. Both are interests. The second at least doesn't bend the ranking of individual products, partly because there is no ranking.

Other people's work, credited as theirs

We aren't the only ones prying at these products, and when somebody else publishes something worth knowing it goes on the radar with their name on it and a link to what they actually wrote.

Those entries are restated in our words rather than quoted at length, same as everything else, and nothing appears without a source you can open. That's enforced rather than intended: an entry with no working link is rejected before it can reach the site. Where a finding is originally somebody else's, saying so costs us nothing and taking credit for it would cost us the only thing this site has.

Corrections to Personal Agent Bench, and where to send them

If something here is wrong, we want to fix it, and a fix lands with a date on it rather than as a silent edit. Each one is logged, with its date, on the corrections page.

Send it to [email protected]. That address is read by the people who write these pages, not by a support queue, and you do not need to be the company we wrote about to use it. Telling us which page and which claim is enough; you do not owe us a polite preamble.

A useful correction identifies the page, the exact claim and a source that supports the replacement. A changed product and an incorrect original statement need different records: one calls for a dated update, the other for a correction. The corrections page records the published changes.

What we commit to. A factual error gets corrected and dated. A finding that has gone stale because the product changed gets re-verified rather than quietly deleted. And a correction from the company whose product we described gets the same treatment as one from anybody else, no more and no less.

The scorecard separates evidence types. The September completion batch comes from the exam team and has no public underlying batch record. The older 114-task comparison was supplied with transcripts that we have not republished. The individual recorded runs show the evidence for specific attempts; they do not make either September aggregate independently reproducible. The 11 October batch, in which each product answered the same 20 tasks once, is the one we checked attempt by attempt against the exam team's records. Public-source product pages identify their sources, and teardown findings describe our observations.

FAQ

Who is behind Personal Agent Bench? The Zentor team. Zentor builds agent software, so the conflict is real and is handled by keeping Zentor's own products off the scorecard permanently rather than by scoring them modestly.

Is this the same as micro1's PersonalAgentBench? No. micro1, a company that supplies expert human data for AI training, published a separate benchmark under the name PersonalAgentBench on its own site. We are not affiliated with it. Its experts scored six assistants on 20 everyday tasks, one run each, and ranked them; this site takes the products apart and reports rates with intervals and no ranking. Other results for the name are unrelated too. MyPCBench is a research benchmark of 184 tasks for computer-use agents on a simulated Linux desktop. SovereignPA-Bench tests whether agents that book and buy for users stay loyal to what the user asked for. OpenClaw's personal agent benchmark pack is a set of local QA scenarios inside that project. Personal Agent Bench is this site, personalagentbench.com.

Why should I trust a bench run by a competitor? Don't trust it. Check it. Teardown findings concern observable behaviour. The September completion batch is not independently checkable because its underlying records are not public; the 11 October batch was checked against the record of every attempt. And the last row of every column tells you what we failed to establish.

Why is there no overall score? There are completion rates from the task runs, printed with their intervals and sample sizes. What there isn't is a single score per product, because combining a completion rate with the runtime findings means weighting them, and the right weights depend on what you're going to use the thing for.

Do you publish the prompts or internal documents you found? No, never. We publish what we concluded in our own words. That covers numbers, structures and behaviours, and it excludes instruction text, internal documents, schemas and any description of how material was obtained.

Will Zentor products ever be added to the scorecard? No. That's the whole mechanism by which the conflict is handled, and adding them later would retroactively make everything published before it dishonest.

How do I report something that's wrong? Email [email protected] with the page and the claim. Corrections get made with the date attached, and a finding that went stale because the product changed gets re-verified rather than quietly deleted.

What Personal Agent Bench is for

It exists so that somebody about to hand an agent their email, their calendar and a payment method can find out how the thing is built before they do it, rather than after.

That's a narrow goal and the site is shaped around it. Start with the scorecard, read how we score if you want to know where the axes came from, or go straight to the four teardowns.