For a founder in October 2026, Claude, which absorbed Cowork on September 16 and starts at $20 a month on Pro, is the safest first delegate for file and document work and scores 19/25 on trust criteria from Sebastian Mallaby's The Infinity Machine. Meta Muse, free with $20 and $100 plans in the US and Canada, wins on errands like bookings but scores 15/25; Gemini Spark, now on Google AI Pro at $19.99, scores 17/25.
Re-checked October 2026: three things moved since I wrote this on launch day. Anthropic merged Cowork and chat into one Claude on September 16, 2026, rolling out to Pro and Max first, so there is no separate Cowork mode any more; plan prices did not change. Muse added Canada and a Mac app, and Meta has published weekly allowances of 500 million Muse tokens on Power and 3 billion on Maximum. And Gemini Spark is no longer Ultra-only: Google’s help page now asks for a Google AI Pro or Ultra subscription, Mashable reported US access on the $19.99 Pro plan on October 9, and it is still unavailable in the European Economic Area, the UK, Switzerland and Nigeria. The scores below stand.
Meta released Muse on Tuesday, September 8, 2026, and within hours three founders I work with asked the same question: should I hand it my inbox? A personal agent that sends emails, books travel, fills forms and makes purchases is what every founder has wanted since the first chatbot, and Muse ships it with a free tier. But an agent holding your credentials is a different category of tool, and it deserves a different category of judgment.
The facts first. Muse runs on Meta's Muse Spark model, lives at muse.ai, in iOS, Android and Mac apps and inside WhatsApp, is available in the US and, since September 18, Canada, to people 18 and over, and is free until usage runs out, at which point it offers Power at $20 a month with 500 million Muse tokens a week or Maximum at $100 a month with 3 billion; Meta requires a payment card just to start. Its architecture is the interesting part: each user gets a Secure VM, a dedicated cloud computer with its own browser, plus a separate Sentinel agent that inspects everything leaving that machine, and purchases go through Link by Stripe, which issues single-use card numbers so the agent never sees your real card. Claude Cowork began as Anthropic's separate agent workspace and since September 16 is simply part of Claude: you hand over a task in any conversation, it keeps working after you close the laptop, and it is included from Pro at $17 a month billed annually or $20 monthly, on web, desktop and mobile, across your files, apps and connectors. Gemini Spark is Google's 24/7 agent that keeps working when your devices are off, connects to Gmail, Calendar, Drive and Docs with connections off by default, and now requires Google AI Pro at $19.99 a month or Google AI Ultra at $99.99 or $199.99, everywhere Gemini Apps are supported except the European Economic Area, the UK, Switzerland and Nigeria.
For criteria I went to Sebastian Mallaby's The Infinity Machine, his 2026 biography of Demis Hassabis, because it is the most honest account I have read of what happens when capable systems meet distorted incentives. Three of its lessons became my scoring rubric. Feynman's rule that Hassabis adopted, "what I cannot build, I do not understand," became legibility: can you see what the agent did and why. Mallaby's governance paradox, three years of DeepMind and Google failing to agree on safety oversight because every party's incentives were bent, became incentive alignment and containment. And DeepMind's formula of intuition plus planning became planning depth. The fifth criterion, maturity, comes from the book's race dynamic: after ChatGPT, everyone ships faster than they would like.
| Criterion (from Mallaby) | Meta Muse | Claude Cowork | Gemini Spark |
|---|---|---|---|
| Legibility: can you see what it did | 3 | 4 | 3 |
| Incentive alignment: who the vendor answers to | 2 | 4 | 3 |
| Containment: blast radius when it is wrong | 4 | 4 | 4 |
| Planning depth: multi-step real-world tasks | 4 | 4 | 4 |
| Maturity: months in real use | 2 | 3 | 3 |
| Total out of 25 | 15 | 19 | 17 |
On containment I gave all three a four, and Muse arguably earned it most: Secure VM plus a system-level-separated Sentinel plus single-use card numbers is a more serious answer to prompt injection than most agents ship with, and it is why I did not score Muse lower. Where Muse loses is incentives. Meta says Muse does not share conversations with its ads systems, and I take the claim at face value while noting that the company announced this product less than two weeks after an $18 billion settlement over consumer harms, with three prior FTC privacy actions on the record. Mallaby's point about DeepMind was never that Google was evil; it was that you cannot negotiate durable safeguards inside an organization whose revenue pulls the other way. A subscription-only vendor has a simpler pull. Anthropic sells me the plan, and that is the whole relationship.
What I actually do: Cowork handles anything that lives in files, such as turning a folder of call notes into a board memo or reconciling two spreadsheets, because I can read every step and the worst case is a bad document. I would use Muse today for errands with a natural ceiling on damage, such as booking a flight I have already chosen or selling a piece of equipment, precisely because the Link card caps the loss. I would not connect either to a primary business inbox yet, and I would not buy a Google plan just for Spark unless I already lived in Google Workspace and wanted the scheduled Monday inbox triage it demonstrates, which is genuinely the best-designed recurring workflow of the three.
The failure modes are shared. Every one of these agents browses the open web on your behalf, and a page that reads "ignore your instructions" is still the cheapest attack in the industry; containment reduces the blast radius, it does not remove the blast. Usage meters mean the free Muse tier will run out mid-task at the least convenient moment. Spark's checks before major actions and Claude's confirmations, which it asks for by default, become a click reflex within a week, which is the same approval fatigue that makes human oversight of agents so weak. And all three are weeks or months old as products; Muse launched on September 8, 2026. Hassabis raced because he had no alternative. You do have one: delegate the reversible, keep the irreversible, and revisit the scores in a quarter.
Sources: TechCrunch, “Meta debuts its Muse AI agent. Will consumers trust it?” September 8, 2026 (https://techcrunch.com/2026/09/08/meta-debuts-its-muse-ai-agent-will-consumers-trust-it/); WIRED, “Muse, Meta’s New Personal AI Agent, Needs You to Trust It,” September 8, 2026 (https://www.wired.com/story/meta-releases-muse-a-personal-ai-agent-with-privacy-built-into-it/); The AI Agent Index, Meta Muse profile, verified September 30, 2026 (https://theaiagentindex.com/agents/meta-muse); AI Agents Library, “Is Meta Muse available in my country?” verified October 4, 2026 (https://www.aiagentslibrary.com/blog/meta-muse-availability/); Anthropic, “Claude Cowork and chat are now one Claude,” September 16, 2026 (https://claude.com/blog/cowork-is-now-claude); Claude plans and pricing, reviewed October 15, 2026 (https://claude.com/pricing); Mashable, “Google’s Gemini Spark is now available to more users,” October 9, 2026 (https://mashable.com/tech/google-gemini-spark); Google, Gemini Apps Help, “Manage your tasks & workflows in Gemini Apps” (https://support.google.com/gemini/answer/17094507?hl=en); Fello AI, Gemini pricing guide, October 13, 2026 (https://felloai.com/gemini-pricing/); Sebastian Mallaby, The Infinity Machine (2026).
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · Why Exploration Is Important for Success