Manus Team Spent $10,000 per Person on AI in September; Engineers Keep a Dozen-Plus Sessions Open at Once

Manus's average AI spend in September exceeded $10,000 per person, with a few colleagues spending $70,000–$80,000 in a single month. Xiao Hong gave this figure while sharing AI Native practices. Where did the money go? The article offers a few more specific slices.

1. Engineers push a dozen-plus tasks at once

In the past, when engineers got a requirement, they first analyzed it themselves, then wrote code. Now requirements and bugs can be handed directly to Codex or Manus, and AI reads the code, locates the problem, and completes the change. Xiao Hong has seen engineers with a dozen or even twenty AI sessions open at the same time. The part engineers personally execute has shrunk; allocating tasks, providing context, checking results, and taking responsibility have become more important.

2. Manus-mono connects code and the R&D workflow

Manus's frontend, backend, and mobile code were originally spread across different repositories. Manus-mono provides a unified entry point and includes an AI-facing guide explaining each repository's purpose, how the repositories work together, and the conventions to follow during development. Through this entry point, AI pulls the code that needs changing, understands the business context, and then uses Git worktree for parallel development.

After that, Manus-mono gradually added local running, online log troubleshooting, configuration reading, release workflows, and end-to-end testing. In the past, pushing code to the test environment required waiting 5–10 minutes; now the full Manus can run locally, and validation takes only seconds. AI can participate in the entire process from understanding requirements and modifying code to testing and release. Key release steps still retain human approval.

3. Product managers go from filing requirements to shipping features

About half of Manus's product managers can independently complete requirements and ship them; all product managers have submitted code to Manus, and some designers have also started writing code.

The article mentions a product manager responsible for a speech-to-text feature. She first had Manus test more than ten models; after choosing a model, she used Mono to get the frontend code, built a demo according to the existing design guidelines, and had Manus generate images for different usage scenarios. She checked the results one by one, and after making revisions, used Manus's video editing capabilities to turn the demo into a video, finally completing the web and desktop features.

Product videos are also beginning to take on part of the role of product requirements documents (PRDs). The video directly shows user scenarios, interface changes, results after clicking, special states, and some implementation logic, making it easier for product, design, engineering, and QA to align.

4. The marketing team starts using Agent colleagues

Marketing partnerships involve outreach, confirmation, script review, pricing approval, contract payment, content publishing, and performance review. In the past, this information was scattered across email, contract platforms, approval systems, project spreadsheets, and various platform backends, and team members constantly moved information between systems.

The Manus team first had AI automatically discover when partnership content and ads were published, then continued to retrieve data, review past partnerships, and connect contract signing, internal approvals, and progress updates. This digital colleague can receive and schedule tasks, manage multiple sessions, judge whether a Slack message needs immediate attention, and remember the team's long-term goals and product information. After the product team releases a new feature, it also proactively learns the feature's applicable conditions and later uses them for external content production. If the internal website it uses lacks a certain function, it can submit a request through the feedback interface and automatically trigger a Manus development task.

5. Sessions begin to become collaborative spaces

In the past, the work process between engineers and AI usually happened in personal sessions, and others who wanted to know the progress had to ask the engineer directly. Manus is trying to let members of the same project view each other's sessions and request to join the conversation. Engineers can first have AI complete a large amount of work, and then a product manager can enter the same session and continue adjusting spacing, styles, and copy. Product, design, engineering, and AI collaborate within the same work process.

Xiao Hong himself is also managing to-dos in a similar way. Manus breaks to-dos into tasks; those that can be executed directly are started, while those that need background information prompt him with questions, then it completes research using content from Slack, Gmail, and Notion, and hands the organized tasks to the corresponding product managers.

6. AI spending buys time and understanding of capability boundaries

Manus's per-capita AI spend in September exceeded $10,000. Xiao Hong believes this money mainly bought the team time, attention, and an understanding of the boundaries of model capabilities. Manus uses the strongest models for code review. After code submissions increased substantially, the code review bot still needs to check critical modules such as payments. The team also uses internal usage data to identify token waste and continuously improves Manus Studio to reduce future costs.

Put together, these changes show that AI Native is no longer just a slogan. Code repositories, product requirements, marketing processes, team collaboration, and to-do management are all beginning to be redesigned around AI. Engineers manage more parallel tasks, product managers drive more complete delivery, and leaders spend more time on judgment and decision-making. AI handles more execution, while people decide what to do, how to accept it, and ultimately take responsibility for the results.

A netizen said in a reply that the line about retaining human approval for key release steps is the most clear-headed. Another said they can no longer tell which comments in the comment section are from AI. This detail is more worth noting than per-capita spending: processes can be handed to AI, but the person who signs off still has to be there.

Original tweet: https://x.com/hocyzon/status/2114562743802175965