The whole point of Mato is to make human-in-the-loop supervision easier, not to encourage autonomous loops.
In my daily workflow I constantly run multiple coding agents at the same time. The annoying part isn’t the AI itself — it’s switching between tabs, terminals, and different tools just to check what each agent is doing.
I built Mato mainly because I wanted a faster way to jump between agents, review their outputs, and approve or intervene when needed. Think of it more like tmux for AI workers, where a human manager can oversee multiple agents at once.
Personally I’m also skeptical of fully self-driving loops. In practice the plan → execute → review cycle with a human in the loop is still the most reliable way to work with AI today.
I built this because most existing agent frameworks felt either too academic (great papers, few real-world tools) or too demo-ish (cool examples, but brittle in production).
We needed something that could actually run GAIA-style tasks end-to-end: reasoning → tool use → verification → retry loops → success.
So GAIA Agent is basically the stack I wished existed:
- Zero-config agent (createGaiaAgent())
- 18+ built-in tools: browser, search, sandbox, memory, filesystem
- Fully TypeScript, modular, and swappable
- Built to run GAIA Benchmark without custom wiring
- Simple enough for side projects, but reliable enough for production
This is very early, but feedback from other agent-devs would help a ton.
If you try it and something feels off, missing, or over-engineered — please tell me.
Would love to hear what kinds of agents you’re building too.
It’s not the AI company’s fault. Think of it like a recruiter: they recommend a cashier, the cashier miscalculates expenses — you don’t sue the recruiter.
Responsibility moves inward, not upward. At some point, it’s about who decided to trust the tool.