What it is
A general agent is a system that takes a task in plain language and carries it through to a result on its own, combining planning, web navigation and tools: from screening CVs to assembling a report. Not a specialized coding agent, but a do-anything executor.
Where it came from
- March 6, 2025 — the Chinese startup Butterfly Effect (Wuhan) launches Manus in invite beta, calling it "the world's first general AI agent".
- Under the hood it orchestrates several models (Claude, a fine-tuned Qwen) and independent agents.
- In parallel: OpenAI Operator and Google Project Mariner — the labs moved the same way.
Why it took off
- The "agent does multi-step work by itself" demo hit a nerve: 1M+ views in 20 hours, with invite codes reselling for thousands of dollars.
- It builds on computer use and browser agents — technically this only became possible in 2025.
- It gave everyone an image of the future "assistant agent for everything" that shifts the whole market's expectations.
How to use it today
- Try it on multi-step routine: research, data gathering, form filling, drafting a report.
- Give it a clear goal and a verifiable output artifact — long autonomous chains are still fragile.
- Keep a human on the final check: the "general" part still runs ahead of reliability.
What to watch out for
- Plenty of noise and inflated expectations — independent reviews (MIT Tech Review and others) also recorded failures on hard tasks.
- Browser access plus data plus outbound connectivity is the lethal trifecta; don't hand private material to untrusted flows.
- A cloud "agent for everything" sends your tasks and data to a third party — weigh the privacy side.