Meta acquires Manus AI for $2 billion — Nate B Jones explains why a "wrapper" commanded that price
Solo analysis by Nate B Jones of AI News & Strategy Daily on Meta's acquisition of Manus.
Summary
Nate B Jones argues that Meta's $2 billion acquisition of Manus is justified not because Manus is a frontier model, but because it has solved one of the hardest problems in AI agents: reliably finishing complex, long-running tasks. He explains that Manus pioneered a set of agentic harness techniques — including KV cache optimization, restorable compression, and persistent file-system memory — that subsequently became best practices across the industry, and that this track record of innovation is what Zuckerberg was really buying. Jones speculates that Meta's primary use case will be an automated ad-building and campaign-management tool, consistent with Meta's stated goal of requiring advertisers to do nothing but supply a budget. He also addresses the difficulty of integrating a small startup's innovations into a large company at scale, giving it less than a 10% chance of succeeding within the year. The episode closes with an assessment of three alternatives to Manus for users who want to move away from Meta-owned tools: Claude Code, Genspark, and Do Anything.
Key Takeaways
FULL TRANSCRIPT
Why Meta paid $2 billion for a "wrapper"
Nate B Jones: Meta just paid over $2 billion for a wrapper named Manus. Not a model, not a breakthrough in reasoning — a wrapper. And ironically, even though I call it a wrapper, I do think it was worth every penny. I think it was worth every penny because of the characteristics of the agentic harness that the Manus team has been able to create.
There are two kinds of agents in the market right now. Most AI agents are really good at starting something. They'll produce a plan. They'll draft an outline. They'll open up tabs. They'll generate a half-done artifact and it looks great — but then they can't finish. Manus has been the flagship for "finish what you start" — for agents that actually do the work — and that is a wide range of work. They are able to do research, coding, data analysis, creating artifacts, building websites.
The important thing — the reason the market cares — is really not that menu of capabilities. It is that they have discovered an interaction pattern that scales generally. You give Manus a goal, it runs a long loop of tool calls, and it comes back with a complete result. That is not as easy as it sounds.
Why the Manus team's transparency strengthened their valuation
Meta clearly felt like the team was on to something special, because they absolutely bought. And you might think: how do you know they're using a special agent harness, or how do you know they're actually good at what they're doing?
Ironically, I think this supported their valuation. The Manus team disclosed a lot of this in a late-summer blog post about how they built long-running agents successfully, and a lot of what they did subsequently became best practice across the community. This is a case where you might think transparency betrayed the secret sauce. But what it really did is show Zuckerberg that this team innovates, this team is able to pay attention to details, and this team can execute against very hard problems.
In their blog post, they talked about things as technical as how often you hit the KV cache in order to reduce the cost and latency of your model. KV cache — key-value cache — basically refers to the idea that you're not going to recompute the input context every time you generate a new token; you're going to hit a cache for it. When you have a very large input string length — a lot of tokens going in, which with agentic artifacts can be a hundred times more information than you're getting out — you want to hit the cache intelligently in order to reduce cost. The Manus team pioneered a lot of different techniques for that.
They also pioneered techniques that keep the agent focused on very long tasks involving lots of different tool calls and lots of working through obstacles. Part of how they do that is by asking the agent to go back and revisit and rearticulate its goals over time. They also pioneered some really interesting work around restorable compression — talking about how you can use a file system as persistent external memory, how the agent can write to that file system, drop things out of the context window, and then come back and recover them.
If I'm Zuckerberg and I'm looking at that, what I see is less "today's solution for today's harness" and more "this team can innovate in the agentic space." They can build really cool and necessary agentic harness innovations that get work done, and the market has clearly shown those innovations are worth purchasing.
What Meta is likely to build with Manus
That's what Meta needs. Meta has had real trouble with their last LLM launch. The LLM itself was reported to have fudged benchmarks — by Yann LeCun, which is publicly embarrassing to Meta. Meta has an entire new team working on the next iteration of Llama, and they need to not fall behind on the usage of the model. That is exactly where Manus shines — as a harness that wraps around a model and makes it useful.
You might wonder where inside Meta, what internal tool, Zuckerberg is thinking of here. My own best guess is ads. That's always how Meta has rung the cash register. That's how they've lived and breathed. My suspicion is they want an automated ad builder — a tool where you sign up, spend a certain amount, and it just builds your ads for you, which you can then execute and run. Meta will optimize your campaigns for you. That aligns with the public declarations the Meta team has made: they really want to be in a position where all you have to do to advertise on Meta platforms is have a wallet. If you can pay, Meta can do all the rest — ad creation, ad spend, ad optimization, campaign creation, segmentation, and so on. Manus is a key part of that.
I will say I don't think it's going to be trivial to take the innovations that Manus created as a tiny startup and integrate them into Meta. If I had to put a probability on that being successfully done this year, I'd put it at less than 10%. It is very, very difficult historically for a large company to take an extremely successful small company, absorb those lessons learned, and scale them in a way that multiplies impact.
We will see. My own guess for now is that Manus will continue to run. It will probably get pulled into Meta's data policies somewhere over the next two or three months — which is why, anecdotally, a lot of people are already moving away from Manus to get ahead of that. In the meantime, Meta will try to launch something around ads, I would guess sometime in the new year.
Three alternatives to Manus
Now, if you've been using Manus, what are your alternatives? It's really tough to find, but here are three, and I'll tell you honestly what I think of each of them.
The first is Claude Code. Claude Code is not primarily a web-based agent. It is becoming very rapidly a general-purpose agent with its roots in the terminal. It's designed first to help you write, edit, and reason about code over extended loops. But because it can now touch your browser through the Claude Code extension, and because it can now touch your files, it can do a lot more than that. People are able to do a lot of finishing work or building work that involves software, but also web research and documents. I saw someone who used Claude Code to get a discount on their Comcast subscription — Claude Code was able to autonomously chat with a Comcast representative and secure the discount for them.
Claude Code is one of those general-purpose tools that we're just now beginning to understand the full width of. It has a very similar harness to Manus in the sense that it's a tool call on a loop. It has absorbed — or independently invented — many of the best practices that Manus talked about in their late-summer blog post: writing good goals into a markdown file, adding skills more recently (which is definitely a Claude Code innovation), and even more recently adding some native built-in eval looping. There was a lot of discussion on Twitter over the Christmas break about the Ralph Wiggum eval loop, which is not super fancy — it's basically just a script that feeds back the prompt and asks the model if it's done until it actually is done.
Claude Code is one option. It does require familiarity with a terminal, and I would say it skews a little more toward code than Manus does.
Genspark is closer to the Manus vibe in the sense that it takes the idea of busy work — documents, slides, sheets, research — and frames that as a problem a super-agent inside a browser can go and tackle. It is probably the closest one-to-one comparison with Manus right now. You'll have to evaluate for your own work whether Genspark is as good at finishing as Manus. Having played with both, I found Manus a little more reliable. But I'm sure the Genspark team has had a fire lit under them post-acquisition, and we may see that shift soon — I would expect them to receive a fair bit of incoming business from people fleeing the Manus acquisition.
The last one I want to highlight is very new — still in alpha. It's called Do Anything, at doanything.com. It is the most directly agentic platform of the three. It explicitly advertises that its agents can connect with 10,000-plus tools over thousands of APIs. The agents have names and email addresses, and during signup it specifically asks you to set a very big and ambitious goal that your agent will work on for a long period of time. Like Manus, it's pay-as-you-go — you pay for the tokens for what you want to get done.
The goaling aspect is a bit different from Manus. Manus has advertised itself as "get this task done completely" — build the slide deck and get it all the way finished. Do Anything is advertising more like "start a business for me" — an order of magnitude or two bigger as a goal. What I found when I played with it is that it still has trouble with the finishing aspects. I gave it a big goal, asked it to start a business for me, and what it came back with was a lot of thinking, a lot of researching, a lot of planning — but then it wanted me to do a lot of the work. The team at Do Anything is still working out the kinks around what full end-to-end execution looks like. They are aiming at a bigger goal than Manus partly because they anticipate gains in model capability that will make that kind of larger goal viable in 2026. They're aiming at where the puck is going.
The bigger picture: harnesses vs. models
Ultimately, I think Meta bought Manus because reliable finishing is a scarce commodity. Think about it this way. Meta may have a team building a model — let's say models are like car engines. They're building a new car engine. You have existing car engines: Claude is a car engine, OpenAI has car engines in Codex, in ChatGPT, and others. But you need a car to go anywhere, and the car is the agent harness. What the Manus team is really good at is building the car.
We are all discovering together what a good agent harness looks like. Claude Code, for what it does, is a very strong agent harness. The Codex agent harness is very good at a different aspect — code quality, code review, and code construction — designed on a slightly different angle than Claude Code, but also very strong in its own way. And we're going to see lots of other agent harnesses coming as well.
My takeaway from all of this is: maybe we should stop thinking about who has the smartest model, and start asking ourselves what it takes — what are the best practices — to build an agent that actually finishes the work it sets out to do. What is the right size of goal for that agent? Do Anything definitely challenges our thinking there. And how do we look at the details the way the Manus team looked at the details when constructing our agentic harnesses, so that we're actually able to build something that does really useful work in a token-efficient, cost-efficient way?
That's harder than it looks. And I think that is the heart of why Zuckerberg decided to go with an acquisition rather than just trying to hire for this. This is not talent that is easy to find in the market at this point.
That's the story on the Manus acquisition. I think this is going to be a year when harnesses are, in a sense, more valuable than the models underneath them. That's one of the big shifts in late 2025 that countered the conventional wisdom of early 2025, when everyone thought the models were going to eat everything and the harnesses would mean less. As we've built agents more and more, the harnesses mean more and more. And that's exactly what we're seeing with this acquisition.