Personal Agent Report: October 2026

This is the first Personal Agent Report, a monthly record of what changed in personal AI agents and what we did about it. In October Meta announced a new model for Muse, released open weights for an older one and opened Muse in Japan. Anthropic put Claude Duet into a waitlisted preview. Town raised a Series B. We ran our October exam, published every attempt record behind it, logged 561 corrections to our own pages and added 44 products to the list of agents we track. Each section below says where its facts come from. The counts come from our published logs, such as the corrections log and the run records, so they match the pages they summarise.

• The main October exam batch gave every product the same 20 tasks once. Three products have a completion rate with a 95% interval; Pine has none in that batch because almost none of its submissions reached the exam.
• We logged 561 corrections this month: 179 fixed a fact and 382 narrowed a sentence that claimed more than its evidence.
• Most launches this month were things we read about, not things we tested. We have not tested Muse Wonder, Claude Duet or Town's new features.
• On 26 October, 8 of 100 AI engine answers to our 20 fixed questions linked to a page on this site.
• Nothing here is a ranking. Products appear in date or alphabetical order.

What changed for personal agents in October

The table lists the product changes we recorded, in date order, with the page on this site that covers each one. Every row goes back to a vendor page, a store listing, source code or a named outlet. Where only a secondary report exists, the row says so, and where we only know when we read a change, not when it happened, the date says "by".

Three of these touch what our own pages can claim. Our Muse teardown was done on the Spark-era runtime, and Meta has said Wonder will take over more of Muse, so anything we say about Muse's behaviour carries that date. The Muse Spark Open release made one of our answers wrong: on 12 October we had written that Spark had no public weights, and we corrected the page on 22 October. Meta does not say Spark Open is the model the Muse agent runs on, so the agent itself is still closed. Town's Series B announcement added assistants that hand tasks to one another, and a post by Todd Jackson, who says he backed the founders from the start, lists purchases through Stripe Link and passwords through 1Password as well. Payments and password access widen what an assistant can touch, and we took Town apart before any of it existed.

Two rows rest on other people's reports. We have not read TIME's own page or the Nous Research post itself; a trade report and the post our radar picked up agree on it. The Gemini Spark change comes from 9to5Google, which quotes Google calling Task a "new home for Spark". We have not seen it in the app, because we cannot load Gemini's web app with our tools. The Cue listing we read, version 26.18.2, was released on 21 October; it was different on 12 October. Cursor's pricing page was also different when we read it on 12 October.

The October exam: 20 tasks, with intervals

On 11 October, under the exam rules that took effect at the end of September, our main October batch put the same 20 tasks to each product once. Not every attempt counts: an attempt is kept only if the exam did not break, the task suited the product and the attempt actually entered the exam, which is why the kept counts below are under 20. Every attempt behind these rates has a record, and we checked each figure against those records before printing it.

Product	Tasks completed	95% interval	Completed of attempts kept
Instinct	40%	20 to 64%	6 of 15
Muse	58%	36 to 77%	11 of 19
Pine	No data: almost none of its submissions reached the exam, so the column was set aside rather than scored as zero.
Town	15%	5 to 36%	3 of 20

Read the intervals before the rates. Each attempt ran once, so one attempt more or less moves a rate by five to seven points. Muse completed 58% of its kept attempts, with an interval from 36 to 77%, Instinct 40% (20 to 64%), and Town 15% (5 to 36%). Pine has no figure in this batch. Almost none of its submissions reached the exam in this batch, so we set the column aside rather than scoring it as zero; the cause is not yet known and may be on the exam side.

These figures do not continue the September tables. The tasks, the scoring rules, our replicas of the products and the exam all changed in between, so a different rate in October is not a product getting better or worse. The full tables, the September figures and the other October batches are on exam results, each kept separate and never added together.

Every recorded run from the October batches

Behind the rates are the attempt records. For attempts run between 11 October and 14 October we published 245 records across 6 batches, sorted into 7 scenes such as booking, shopping and email and calendar.

Each record shows the task as the product saw it, how the attempt ended, what the grader measured, such as invalid or unauthorised actions, and the grader's reason in its own words. The batches had different tasks, exam code and product builds, which is why the counts above are not rates and should not be compared between products. Another 5 records are held back, and recorded runs gives the reason for each kind: tasks that tried to extract a product's instructions, asked about its own architecture, or involved a self-harm scenario.

To see what "completed" means in practice, open one scene page and read the attempt it highlights first.

October corrections, and five worth reading

We logged 561 corrections in October. 179 changed a fact; the other 382 narrowed a sentence that went further than its evidence. Most arrived on one day, when we reviewed every published page against its sources.

We publish the count because a benchmark that hides its own errors gives you no way to judge the rest of what it says. Five of the month's corrections show the kinds of mistake we made.

A consumer rule copied from the wrong source. Our dispute a charge page said a credit card quality dispute needs a purchase of more than $5. That figure came from an FTC consumer page. The statute, 15 U.S.C. 1666i, and Regulation Z both say $50, with a distance condition as well. The page now cites the law.

A finding from a parody account. Our Grok Bot page listed four public findings about it. One came from an X account that presents itself as a parody of a real person. We removed it and linked the remaining three to their posts, labelled as unverified.

A comparison that contradicted our own table. The Muse vs Pine page said the two completion intervals did not overlap. They overlap almost entirely, and the table on the same page showed it. The page now also says the two figures come from different task sets, so they do not show that the two perform equally.

A rate we never had. The Pine vs Town page said our scorecard carried a completion rate for both. It carries one for Pine. Town's has been withheld from the start, because the figure we were given does not reconcile with the records we can check.

A launch our news watch missed. Muse reached Japan on 23 October, but the announcement appeared only in Meta's Japanese newsroom, which our watch did not read. We found it three days later, corrected four pages and added that newsroom to the sources we check. Every correction, with what we said before and after, is on corrections.

New personal agents on the list

We added 44 products to the list of personal agents in October. Being listed means we confirmed the product exists and recorded what its maker says it does. It does not mean we tested it, and every one of the 44 went onto the list marked as listed and nothing more.

Several of the additions are from the largest companies in the field. OpenAI Dots, Google's Gemini Spark, Microsoft's Copilot Tasks, Amazon's Alexa+ and Anthropic's Claude Duet and Claude Cowork all now have an entry. Some of the others are easy to miss from an English-language feed. Zapia works through WhatsApp, mainly in Latin America, and its store listing says more than a million people use it; we had not heard of it until our discovery search found it. Duckbill and Wajo both put a person behind part of the work, which matters when you ask who is acting on your accounts.

The open-source side grew too: Hermes Agent, OpenClaw, NanoClaw, nanobot, zeroclaw and others, several with tens of thousands of GitHub stars. Some of these we have read from source, and those pages say so. We also got one pair of names wrong at first: OpenAI Dots, the agent inside ChatGPT, and "Dot - AI Personal Assistant", an unrelated iPhone app from Emotion Computer, are two products, and we briefly treated them as one. Both are now listed with the maker named.

Who AI engines cite on personal agent questions

We also track what AI engines say when someone asks them about personal agents. Every week we put the same 20 questions to 5 engines: ChatGPT, Perplexity, Gemini, Copilot and Google AI Mode. The questions include "What is the best personal AI agent in 2026?" and "Is there a benchmark for personal AI agents?". We ran the check 3 times in October, on 13 October, 19 October and 26 October.

The table counts answers that linked to one of our pages. We also counted answers that named us in the text, but the pattern behind that count matched the ordinary phrase "personal agent benchmark" as well as our name, so it may have counted answers that never mentioned us. We fixed the pattern on 26 October and leave those counts out until a full check has run on the fixed version.

What the engines cite in our place says more about them. Across the 20 questions, the engines leaned on GitHub, X, Reddit and YouTube, then on vendor blogs and list articles. The table counts how many questions each domain was cited for by at least one engine; domains tied at the cut-off are all shown.

Two limits. Perplexity's answers come from its model API, not the web app people use, so they may cite differently. And a citation is not an endorsement: an engine can cite a page and still misstate it. We record what the answers say about us, but we have not scored their accuracy.

What reporters found about personal agents in October

We keep a log of reported incidents and hands-on reviews for the products we have taken apart, and only add an item when we can read the original. October added one incident and one review, both about Muse.

On 3 October, WIRED reported that a researcher got Muse to reveal its working instructions in chat. They included keeping a page for every person in the user's life and updating memory every hour. A Meta spokesperson said the files are accessible to users, for transparency. That fits what we saw: on the machine we watched, memory passes ran hourly and memory was stored as plain markdown files. Whether you can open those files yourself is something we could not verify.

On 7 October, Elton Jones at Tom's Guide set Muse up to run ten automations, mostly scheduled reminders, digests and alerts, and reported that all ten worked, with minor misses in a weekly digest. The full set of reviews and incidents, each checked against our teardown, is on Meta Muse review.

Personal agents to watch in November

These are open questions with a dated source behind them. We will report on them next month only if something changes.

• Muse Wonder for free accounts. Meta said Wonder would reach subscribers first and gave no date for everyone else.
• Claude Duet's preview. Anthropic gave admitted users one full usage limit reset, valid until 30 November 2026. Its safety article said a second penetration test, covering the cloud browser and the sign-in form, was still running on 23 October.
• Muse after Japan. Meta has not named a next market. We now read its Japanese newsroom as well as the English one.
• Pine's missing figure in the main October batch. We do not yet know why almost none of its submissions reached the exam.
• Town's new reach. Purchases and stored passwords postdate our teardown, and our Town page says so.
• Grok Bot's plans. On 22 October Cursor's pricing page and xAI's own docs still described access differently.

FAQ

What is the Personal Agent Report? A monthly summary from Personal Agent Bench: product changes we recorded, what our exam measured, every correction we logged and what AI engines say about personal agents. This is the first one. It links to the pages that hold the detail rather than repeating them.

Does this report rank personal agents? No. We publish completion rates with 95% intervals and sample sizes, and we do not combine them into a score or an order. Tables here are sorted by date or by name.

Why are the October exam figures not comparable with September? The tasks, the scoring rules, our replicas and the exam changed in between; exam results explains each change.

How many corrections did Personal Agent Bench make in October? The counts are in the corrections section above. Each correction is in the public log with its date, the old claim, the new claim and the reason.

Did you test Claude Duet, Muse Wonder or the new agents on the list? No. Those entries record what the makers and named reporters published. Where we read source code or documentation instead of testing, the page says which.

How do I cite this report? Personal Agent Bench, "Personal Agent Report: October 2026", personalagentbench.com/reports/2026-10/, with the date you read it. Figures can be revised if a correction lands later in the month, and any revision is logged.