At Manus, Nobody Has to Explain Why AI Matters Anymore

Yu Yi left the Tencent Research Institute and, within three days, completed her Tencent offboarding and Manus onboarding. On the podcast AI Alchemy, she talked about what the new company feels like, with very direct emotion: so happy, so happy, so happy.

The really interesting thing is not the emotion, but the way of working. When recommending this interview, Zhang Lixing mentioned a phrase: agents are first-class citizens inside the organization. Context can be obtained easily, without people manually moving it around. That sentence is more concrete than a lot of AI-native organization marketing.

First, No Need to Persuade Others to Use AI

Yu Yi says the happiest thing at Manus is that everyone is a heavy AI user. She no longer has to explain to others that AI is good, what to watch out for with AI, or how humans and AI should actually collaborate. At Manus these are common sense. Everyone has their own agent, or their own harness.

In big companies, the divide over AI is severe; the more people there are, the more severe the divide. At Manus, you cannot find anyone opposed to AI.

This sounds like a recruiting ad, but in daily work it is very concrete. She raises a request, and twenty minutes later someone asks her to test it and asks whether it is what she wanted. The team has more than a hundred people; a colleague on the main product team had their phone battery drop from 90% to 5% and still could not get back to their desk. Messy vitality says more about the situation than neatly organized reporting materials.

Onboarding Starts by Chatting with a Bot About Organizational Context

A lot of Yu Yi's onboarding context came from chatting with bots in the organization: what meetings are happening recently, progress on Manus 2.0 and Cue, organizational relationships, and the management style and communication style of Xiaohong and Tao. When she needs to find a particular person, she also first asks AI who is suitable.

After she connected Slack to Cue, when scheduling a meeting it would automatically pull up 'what I recently discussed with Tao and what still needs progress,' and search across channels by person's name to find past content and the last discussion of the topic. Her reflection was that being a secretary really still requires context.

In Manus's Slack, it is not a hundred people throwing a hundred bots into it. In the post Zhang Lixing mentioned, Yu Yi confirmed that the organization has only a few regular bots, centered on feedback, Cue, and general Manus business services. Permission to create bots is in the hands of a few people. People are still more important; AI is used to reduce friction and burden.

Talk to Only One Agent

Yu Yi now chats with only one main agent, called 'AI Yu Yi.' She does not chat with other agents. The main agent is not responsible for doing the work; it is responsible for allocating, monitoring, and accepting it. She even reminds it to distribute the work out; whenever it wants to do things itself, she scolds it.

The key to this approach is the degree of delegation. She quoted a line from an interview: 90% delegation and 100% delegation are worlds apart. With Claude Code or Codex, people still have to watch for a while, keep an eye on it, and in the end there are still blockers. Once the main agent accepts the work, the matter is done; the person does not need to open the process again.

She set up a monitoring bot for each source—email, X, fundraising information, etc.—which summarizes to the main agent for judgment, then hands it to a researcher bot that calls local Codex for deep research. She also delegated the intern interview scheduling process: scan emails, review résumés, check calendars, send invitations, and create Google Meeting links. During the half-day of account migration, she had to schedule interviews herself and found it extremely painful.

Roles Widen; Product and Operations People Also Write Code

Inside Manus, operations, customer service, and marketing can also jump straight into code, which after standardization is merged and deployed. Yu Yi confirms this happens a lot. Everyone's scope of responsibility is much larger than before, and much broader. On the product side, everyone has started writing their own code, then having a developer review it once before merging and deploying. People who have not written code for years are writing it too; the boundary between product and development is extremely blurry.

This is close to what Boris Cherny called role melting in that tweet. Engineering, product, design, and DS blend into new roles, and the team needs different types of prototypers, builders, cleaners, growers, and maintainers. Yu Yi's version is more direct: build it first, and if it feels wrong, cancel it. She has experienced twice being stopped halfway to change direction; no one argued, they just dropped it and redid it.

Delete a Year of Harness, and Prompts Return to One Sentence

Yu Yi plans to delete all the harnesses she has accumulated over a year, rerun everything with the best model, and let AI prove which ones should stay. The trigger was making a video with Manus 2.0: a single prompt produced a very cool effect, while adding her own skill actually made trouble for the model. She judges that a lot of accumulated work has become debt.

Prompts return to one sentence: what counts as good, what counts as acceptable, and what is the essence of this matter. She gave an example of a special effect: AI kept failing to make a PPT effect that coordinated with the narration, and the final sentence was 'all special effects must coordinate with the content and emotion of the narration.' Give AI the why and the what; leave the how to it.

Hiring has changed too. First look at the work—videos, games, small web coding projects are all fine. Then ask what they have built with AI and what they did in side projects. Suppose you came to Manus with unlimited tokens and with product and development support behind you: what would you want to do? Write it out. Good candidates are directly given a batch of Manus credits and asked to build something in a day or two to show. Educational background is not the focus. The work speaks, and the process of collaborating with AI speaks too.

There is also a line that netizens singled out: know-how all belongs to AI; we use know-who to solve problems. We both know many know-who friends. After AI flattens execution, trust relationships and scheduling ability become more expensive instead.

The threshold for an AI-native organization may not lie in the number of models or the list of tools. Whether context flows by default, whether human bandwidth is taken seriously, and whether hiring can reveal how a person collaborates with AI—these are harder. Manus's approach may not suit every company, but at least it puts a question on the table: if you still have to spend effort persuading colleagues to use AI, you are still some distance from AI-native.

Original podcast link: https://www.xiaoyuzhoufm.com/episode/6ae1d8b397726cd4675926fa