The Manus Breakthrough Nobody Saw Coming

Why multi-agent systems will win

If you've been following AI news over the past couple of weeks, you've undoubtedly heard about Manus.

The viral AI agent from a little-known Chinese startup (now being called the next DeepSeek) is a masterful display of a reasoning LLM using tools.

And candidly, watching it solve complex problems by seamlessly reasoning through multiple steps left many people (including myself) genuinely impressed.

But here's the fascinating part: It was only after the agent went viral that we discovered what powered this breakthrough. Turns out, Manus is under the hood is Claude 3.5 Sonnet with access to 29 carefully selected tools.

In today's newsletter, I'll break down why this matters for the AI agent landscape and why I'm more excited than ever about what we're building with agent.ai.

—Dharmesh

What Makes Manus So Special?

Manus immediately captured everyone’s attention when it launched. The world was genuinely shocked by how powerful it was, for good reason.

What we witnessed was a masterful display of what happens when you combine a reasoning LLM with the right tools.

In the viral demos, you can watch as it jumps between tools — grabbing information from one source, analyzing it with another source, and presenting coherent results — all without stumbling through the handoffs that typically break AI workflows.

Then came the surprising revelation: Manus is "just" Claude 3.5 Sonnet with access to 29 carefully selected tools. That "just" is doing a lot of heavy lifting, because the combination creates something far greater than the sum of its parts.

But in the end, the real breakthrough wasn’t the model itself (though Claude 3.5 is certainly impressive), but rather what happens when you:

  1. Give a reasoning-capable LLM access to a diverse set of tools

  2. Let it figure out which tools to use when

  3. Allow it to chain operations together into complex workflows

While Manus is currently only accessible via invite code (which they've stopped distributing due to server limitations), what makes it truly special isn't its exclusivity, it's what it reveals about the future of AI agents and what I like to call Multi-agent systems.

This is exactly what we've been building toward at agent.ai, but with an important difference: scale. More on that below.

Our Approach to Multi-Agent Systems

In case you couldn’t tell, I'm pretty excited about agent.ai's potential to take the concept that Manus exposed to the world exponentially further.

While Manus uses a single LLM (Claude 3.5) with 29 tools, we’ve been heads down building a platform that:

  1. Leverages multiple frontier LLMs from OpenAI, Anthropic, Google, and others

  2. Has access to 1,000+ community-built agents — each of which can function as a specialized "tool"

  3. Creates an ecosystem where builders can build on top of each other's work

Imagine what happens when thousands of builders create specialized agents that become tools for other agents to use. Each builder focuses on their area of expertise, creating increasingly sophisticated capabilities that combine in ways we can't even predict yet.

The magic happens when we connect the dots — when the tools built by thousands of specialized builders become available to multiple frontier LLMs.

And this is all happening at a perfect moment, as we're witnessing the simultaneous evolution of three critical components:

  1. Smarter Models: Each new LLM release (like Claude 3.5, GPT-4.5) brings better reasoning capabilities

  2. More Tools/Agents: The ecosystem of specialized tools and agents on agent.ai is growing rapidly

  3. Interoperability Standards: MCP (Model Context Protocol) is gaining traction as an open standard for how tools can be made available to LLMs (we already have a beta of MCP server support for agent.ai)

This three-part evolution creates a compounding effect.

Better models can use tools more effectively. More sophisticated tools give models greater capabilities. And standards like MCP enable everything to work together seamlessly in robust multi-agent systems.

The Multi-Agent Future

So as impressive as Manus is, it's just the beginning.

The future is multi-agent systems where specialized agents collaborate, sharing capabilities and building on each other's strengths.

That's partly what agent.ai is about — creating an ecosystem of agent builders whose work can be used as tools by others.

Manus demonstrates what's possible with one model and 29 tools, but agent.ai aims to unlock what happens when thousands of specialized agents become available to multiple frontier models.

This is also why the community aspect of agent.ai is so crucial. We're building a network where thousands of builders can collaborate, combine their expertise, and create capabilities that would be impossible to build in isolation.

We still have lots of work to do, but that's the vision we're working toward. And seeing breakthroughs like Manus only reinforces that we're on the right track.

In sum, the magic happens when we connect the dots: Smarter models + more/better tools + standards for agents/tools/LLMs to interoperate.

The future is multi-agent systems.

—Dharmesh (@dharmesh)

What'd you think of today's email?

Click below to let me know.

Login or Subscribe to participate in polls.