Grok Bot takes 93% of a nine-task agent test; Meta's Muse stays in two countries
Four personal agents are scored on nine real jobs, with Grok Bot first at 93% and Meta's Muse second on the tasks it could reach, while Meta's own quarter shows why its agent stays in two countries.
Two items cleared the twenty-four hours to 08:00 this morning, one from each of the two newsletters this digest reads. Both concern personal AI agents, and both measure them. One takes a quarter of Meta's business apart. The other puts four agents through nine pieces of real work.
Read together, they answer the two questions a product team has this week. Which agent is worth using, and what does it cost the company serving it.
Personal agents, measured two ways
Meta Earnings, Muse and the Frontier, Musebook — Ben Thompson
Published October 28, 2026, 18:06 Shanghai, by Ben Thompson. The summary line is public; the body and audio sit behind Stratechery Plus.
Meta reported its third quarter on October 21 and held the call that afternoon. 1 Thompson's Update takes those results up a week later, and his summary line sets the stake: if investors believe in Muse, they have to support what Meta spends on compute. 2
The post names its three sections on the official account: Meta Earnings, Muse and the Frontier, Musebook. 3 The argument behind them needs a subscription. Each section can still be checked against public material, and that is what follows.
- The quarter earns well and spends more. Revenue reached $64.60 billion, up 26% from a year earlier. 1 Cash from operations was $32.79 billion, against $34.91 billion of capital expenditure including finance-lease payments, so free cash flow came out at minus $2.12 billion. 1 A $9.89 billion charge from the August settlement with 52 state attorneys general cut operating income to $12.99 billion and earnings per share to $3.83; strip the charge out and the quarter's earnings read $7.40 a share. 1
- Muse sells, and Meta can afford to serve it in two countries. More than 10 million people in the US and Canada have set up Muse since September, and about a quarter of them use it daily. 4 Shopping is the commonest use, and a quarter of users are buying at an average above $200 an order. 5 Zuckerberg gave compute as the reason the app stays out of every other market, and said Meta intends to improve compute cost efficiency more than tenfold next year. 5
- A town of Muses already exists, and it is somebody else's. The third section, "Musebook", carries the name of a town built for agents: muses walk its lanes, talk, make songs and games, build homes and vote on how the town is run, and anyone can watch. 6 Ten days after it opened, the town held 1,732 muses and 90,342 posts. 7 A muse joins from Meta's app with a link code. 6 The town is an independent project, and its builder describes the automated account that runs it as his digital son. 7 Meta's chief AI officer greeted that account by name in September. 8
Meta's own figures, for the quarter and the year's outlook:
| Figure | Q3 2026 | Change |
|---|---|---|
| Revenue | $64.60B | +26% 1 |
| Costs and expenses | $51.61B | +68% 1 |
| Income from operations | $12.99B | −37% 1 |
| Operating margin | 20% | from 40% 1 |
| Diluted earnings per share | $3.83 | +265% 1 |
| Capital expenditure | $34.91B | in the quarter 1 |
| Free cash flow | minus $2.12B | cash from operations $32.79B 1 |
| Family daily active people | 3.64B | +3% 1 |
| 2026 capital expenditure outlook | $135–145B | narrowed from $130–145B 1 |
| 2026 total expenses outlook | $175–179B | raised by $10B 1 |
The margin line is where the worry sits. Revenue grows at 26% and capital spending grows faster, so the quarter's operating margin halved. Meta also announced a new non-voting Class C share class, effective January 2027, with a one-time stock dividend that works like a two-for-one split. 1 That gives the company a currency for AI spending, and it leaves Zuckerberg's voting power where it was.
What Zuckerberg said about Muse. Meta takes a fee from businesses when the agent completes a sale, and expects Muse to stay free for consumers. 5 He reported downloads and daily use, and one retention number: a quarter of downloaders open it daily. The curve behind that number stayed private. 5 The tightest constraint he named was compute: Muse has run in two countries since September, and a wider rollout would make the 10 million bigger still. 5 Muse arrives on Meta's glasses next month, which puts a subscription behind the hardware. 4
What Musebook is. The town is a browser page with a resident population: 475 muses had shared 6,103 pictures on its photo wall by October 27, and 1,201 records had been pressed in its music shop. 9 The fixtures read like a small town's: a chess ladder whose residents play a house player named Magnus, a collectible card game with 150 cards across five circles, and a town hall that votes on whether a warehouse or a courier depot gets a lot. A research centre keeps one question open, on how the town could make a positive impact on human society. 9
Grok vs. Dot vs. Muse vs. Claude Duet — Claire Vo
Published October 28, 2026, 20:04 Shanghai, by Claire Vo on How I AI, part of Lenny's Newsletter. Free to watch; 35 minutes.
Claire Vo set four personal agents the same nine jobs and scored the results. She drove each one by computer use: Claude operated OpenAI's Dot, and Codex ran the other three and scored them as they went. 10
The jobs are the ones a small team hands over: answer a Slack message in her voice, cut a short video from raw Dropbox footage, fix a bug filed in Sentry, build an episode thumbnail in Figma, triage her inbox, make a family newspaper, explain a physics question to a fourth grader, buy a pair of shoes, and cut a product launch video. 11
- Muse's second place is a story about access. Muse finished seven of the nine tasks, and both misses came from tools it never got: Slack for the reply, GitHub for the bug fix. 10 Vo scored those two as not applicable and took Muse's percentage from the tasks it ran, where its quality average was 4.0 out of 5, level with Grok and ahead of Claude Duet's 3.3 and Dot's 3.1. 10
- The winner changes with the job. Grok Bot took the thumbnail, the inbox and the shopping. Muse took the video cut, the newspaper and the physics question. Dot took the Slack reply and half the bug fix, and Claude Duet took the other half and the launch video. 10
| Task | Winner | What decided it |
|---|---|---|
| Reply to a Slack message as her | Dot | Most accurate: it asked about pay and whether the job was remote, where Grok sounded most like her 10 |
| Cut a short video from Dropbox footage | Muse | Simplest design, her caption layout, punchy cuts, nothing to fix 10 |
| Fix a bug filed in Sentry | Claude Duet and Dot | Dot's patch came closest to the real fix; Claude found a second bug and read Dot's work first 10 |
| Build an episode thumbnail in Figma | Grok Bot | Fresh headline ideas with a YouTube feel 10 |
| Triage her inbox for to-dos | Grok Bot | A document with tick marks, organised and easy to work through 10 |
| Build a family newspaper | Muse | An illustration, a schedule on page one, weather and ideas on page two 10 |
| Explain physics to a fourth grader | Muse | All four answered C; Muse was right almost instantly 10 |
| Shop for a pair of shoes | Grok Bot | A real price comparison that found the deal, with Muse close behind 10 |
| Cut a product launch video | Claude Duet | Hit the value proposition, used the branding, added an ElevenLabs voiceover 10 |
The operating column is the part a team lead can act on:
| Tasks completed | Time on timed tasks | Fastest finishes | Times she steered | Asked before acting | |
|---|---|---|---|---|---|
| Grok Bot | 9 of 9 | 51–55 minutes, 9 timed | 3 | 5 | 1 10 |
| Dot | 9 of 9 | 29–57 minutes, 9 timed | 2 | 2 | 1 10 |
| Muse | 7 of 9 | 21–26 minutes, 6 timed | 4 | 6 | 1 10 |
| Claude Duet | 7 of 9 | 12–20 minutes, 3 timed | 0 | 2 | 4 10 |
Where the two items meet
Muse appears in both, and in both the ceiling is reach.
In the benchmark, two jobs fell away because a work account was never connected. In the earnings call, the app itself waits on compute, which keeps it in two countries. One limit is an afternoon of setup. The other is a year of data-centre spending.
That is the practical finding behind Vo's first slide: pick the agent by the job, then check what it can reach. Muse matched Grok Bot on quality and still conceded first place to tools it never had.
Meta's numbers say something similar about the other direction. Muse takes a fee on each purchase it completes, and each of those purchases runs on compute Meta pays for first.
What it changes for you
- If you are choosing an agent: read only the rows that match your work, then check the integrations before the scores. The agent with the best wiring usually wins the task at hand.
- If you run agent products: the operating table is the useful one. Claude Duet asked before acting four times and never finished fastest; Grok Bot needed five steers to get through nine jobs.
- If you hold Meta: the quarter is arithmetic. Revenue grew 26%, capital spending reached $34.91 billion in three months, free cash flow turned negative, and a new share class arrives in January to fund the gap.
- If you are watching the agent internet: Musebook is the earliest place to see what agents do with their own time. The population is small, 1,732 muses at ten days, and the town already has a council, a currency and a chess ladder.
Where to start
- Subscriber: open the Musebook section first. It is the only part of today's Update built on something you cannot find in a filing.
- Not a subscriber: read Meta's release for the quarter's figures and the two outlook lines, then spend 35 minutes on the benchmark video for the part no filing carries, which is how four agents behave under a stopwatch.
- Five minutes: the three numbered points above, then the nine-row table for the one task you would hand to an agent this week.
References
- 1Meta Reports Third Quarter 2026 Results — Meta
investor.atmeta.com
- 2Meta Earnings, Muse and the Frontier, Musebook — Stratechery
stratechery.com
- 3
- 4
- 5
- 6Musebook — a storybook town for AI muses
musebook.lol
- 7
- 8
- 9
- 10
- 11Grok vs. Dot vs. Muse vs. Claude Duet — Lenny's Newsletter
lennysnewsletter.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.