The Future of General AI Agents: The Story of Manus AI - SuperAI Singapore 2025 (AI column)
Looked at the old one. video Tao Cheung, co-founder Manus AI, shared how the company has gone. Of course, there is no final stage - the acquistion of the company banned in Russia Meta (Since the video has been over six months). In fact, Tao tells how the product went from a browser extension, an experiment with an AI browser, to a cloud agent that does things, not just answers questions. In this talk, there are few words about technology and a lot about how to make AI useful in real-world scenarios.
The highlights of this 24-a minute video like that.
1General AI agent = brain + hands The key metaphor is that one generation of text is not enough. An agent needs “hands” – the ability to perform actions. (Tools/Integrations/Targets)Otherwise, it remains a smart chat assistant. By the way, the name of the startup from Latin translates as a hand. (brush).
2ые Path: extension → AI-browser → cloud tasks The team began by expanding the browser as a way to “embed AI where the user is already working.” Then there was the call to the “AI-browser”, on which the team worked 7- months, but he was uncomfortable and broke the usual flow (I had to sit and wait for the agent in the browser to work and not interfere with it.). This experience led to the idea of cloud-based task execution, where the agent does the work in the background and gives the finished result.
3AI should live in a work context, not in a separate tab Expansion as a wedge product is a bet on “AI in the workflow” (browser/pages/content)It's not a separate chat room where you have to drag and drop the context every time. They were inspired by how Cursor works. (IDE for developers)
4The Most Frequent Yuzcays (And why is it important?) In Q&A, Tao talked about three main scenarios: collecting/researching information, creating slides, assembling sites. This is a good marker of where agents are giving value right now:
- a lot of routine,
- a lot of "gluing" sources / artifacts,
- clear result (Ready-made deck/page/report).([SuperAI Singapore][4])
For engineers, the following ideas can be found.
1) The paradigm is changing: you are not building a chat, but a task execution system. An agent is a mini-distributed system: planning → steps → tools → intermediate states → final artifact → logging. Even if the brain is an LLM, the engineering part is orchestration and control. 2) Architectural Pattern: Sandbox + Connectors Manus clearly has a model: Cloud sandbox/browser for isolation and repeatability Plus local connector/extension where authorized sessions and trusted context are important (logins, IP, tabs) For your internal agents, this translates into: where to perform (locally/in the cloud)with what rights (least privilege)how to audit (log) 3) Context Engineering is More Important than Prompt Engineering If the agent collects slides/pages/reports, it is not the beauty of the prompt that decides, but:
- which sources are connected, What input/output formats are standardized
- how the agent receives feedback and corrects.
This is what it means for technical managers.
1) You need to choose scenarios for automation, and not “paint AI” with an even layer The most “quick wins” are processes with clear Definition of Done and low blast radius.
- preparation of the report/resort, generating a presentation from RFC/PRD,
- assembly of the landing page / demo page, Automation of repeated web operations 2) KPI for agent = speed × quality × price Enter metrics, for example.
- % of tasks "delivered without alterations",
- average time to result, cost (money/credit/tokens), Number of human interventions 3) Risk management: access, audit, shutdown If an agent “clicks and logs in,” you need politicians: Who can connect which integrations,
- where secrets are kept, What does the audit trail look like? How quickly to “stop execution”
In general, the guys from Manus did so well that they were bought by Meta.
#AI #Engineering #Software #Management #Leadership #Startup #LLM #ML #Architecture