Ji Yichao on the Journey from Browsers to Manus and Agent Products

Zhang Xiaojun interviewed Manus co-founder Ji Yichao on December 1, 2025. The episode description says Meta announced an acquisition afterward; the conversation itself was recorded before that event.

Ji recalls developing a browser and apps early in his career, when overseas payment channels, domestic user feedback and the startup environment shaped his choices.

He later became interested in natural language processing, and his original browser gradually ceased being maintained as operating systems changed. He sees its international users and paid sales as an important beginning.

He recalls early models and a project called Maggie, in which the team tried to read web pages and organize knowledge while confronting limited context lengths.

Reflecting on a previous venture, he says he prefers technical and product work to a CEO’s commercialization and personnel-management responsibilities, so he sought complementary partners.

He discusses searching for an AI product direction and realizing that product teams cannot merely wait for model improvements; they must identify what users need.

In his view, model iteration is uncertain. One possible path is to find product demand and a relatively stable experience first, then consider model training to improve reliability or costs.

He says startups face repeated setbacks, and teams need physical and mental resilience and calm judgment to adjust after failures.

The team once tried moving from the Monica browser extension to a standalone browser, hoping to escape distribution and device-side constraints.

They later reconsidered because switching away from a mature browser is costly for users; new features alone may not change habits.

He recalls the team’s move toward Manus and shares its account of product growth. Revenue figures mentioned in the episode are company claims, not independently verified in the interview.

The team observed non-engineers using coding tools for writing and data analysis. They reasoned that code could be an intermediate tool for tasks without being the primary interface users see.

He wants agents to run asynchronously in the cloud so users need not wait continuously and can schedule multiple tasks. Actual experience still depends on quality, speed and reliability.

He distinguishes chat products from agent systems: in addition to the user and model, an agent needs an environment, tools and ways to interact with the outside world.

He uses long-tail needs to explain the value of general tools: some users must process unusual files or complete multistep work that a single fixed feature may not cover.

He sees the user, model and runtime as system boundaries and says the team provides isolated task sandboxes. Isolation still requires specific security design and validation.

On cost, he says multistep agent tasks may repeatedly read context, so token use and cost can differ from simple chat. Ratios depend on the task and implementation.

He discusses context compression in long tasks: a system must preserve important state while signaling that part of the history has been summarized rather than simply disappearing.

He argues that reasoning optimized for mathematics or competitions may not directly suit every agent task; combinations of action, observation and timely reasoning need practical evaluation.

He discusses a remote-work benchmark and values whether systems can complete tasks customers would pay for, while acknowledging results were still far from comprehensive reliability.

He describes users correcting agents and adjusting results, arguing that feedback can improve the system; its collection and use should consider consent and privacy.

He recalls targeting overseas users at launch and responds to questions about publicity. These are his and the company’s positions.

He compares agents’ ongoing compute costs with traditional low-marginal-cost software, saying companies must balance task quality, pricing and long-term sustainability.

On positioning, he says some users will pay for higher quality and may use multiple AI tools. This reflects his business judgment at the time.

Using chart-rendering errors as an example, he says a team might let a system inspect images for a broader range of problems instead of patching only one fixed error, though self-checking also needs validation.

He says building an agent involves both a product for people and a tool environment for the model, connected through technical architecture.

He describes work across sandbox infrastructure, agent engineering, evaluation and research to support system operation and iteration.

He says the team does have researchers but relies more on external foundation models, focusing internal research on context, tools and system questions.

He distinguishes online learning through parameter updates, personalization and continual learning, noting that the terms address different problems.

Looking back on Manus’s year, he says global competition requires continual learning and that rapid growth brings compute spending and organizational complexity.

He imagines asynchronous agents may change the relationship between user attention and product use, while admitting the long-term constraints remain unclear.

Finally, he discusses research directions such as latent reasoning as possibilities worth exploring, not problems already solved.