Context Not Control: The Message Netflix Wrote for Employees, Agents Have Already Put It to Use

The Manus team recently published “Transforming Manus with Manus: Practice and Reflections on Rebuilding an AI-Native Organization.” The article includes a scene: on an engineer’s computer, a row of Codex and Manus sessions is open; clicking into one shows the input is the original text of a requirement just posted on Slack, unchanged.

In the past, when engineers received a requirement, they first analyzed it themselves and then started writing code. Now new requirements and new bugs are mostly handed directly to AI. The AI finds the relevant code on its own and makes the changes itself. The part people execute by hand is getting smaller and smaller, while the share of directing AI to do the work and taking responsibility for the results is growing. The Manus team mentions that the typical pace for engineers is one person handling a dozen or twenty requirements at the same time.

In a tweet, Ma Tianyi summed up this feeling in one phrase: Context Not Control. In the past, people directed AI’s work and used strict constraints to make the agent copy what was expected. Now that model capabilities have become stronger, we should let go and allow agents to explore, rather than let human cognition limit the agent. Effort should go into optimizing the R&D processes and toolchains that agents will use, and into connecting materials and information to agents. People mainly provide information at key nodes, key choices, and key constraints.

A commenter added the source in a reply: “Context, not Control” comes from Netflix culture. ByteDance’s organizational culture drew heavily on Netflix. Ma Tianyi said he had bought the book on Netflix culture long ago and was deeply struck when he read it, but later forgot where the phrase came from.

The Manus team created a code repository called Mono. The code was not merged into one place. Mono provides a unified entry point, along with a manual written for AI: what each repository does, how they work together, and what standards R&D follows. When AI needs to change a particular part, it pulls the corresponding repository.

Mono grew step by step, with each step solving the problems exposed by the previous one: letting AI see the whole picture, filling in business knowledge beyond individual repositories, and supporting cross-repository collaboration and parallel development; letting AI verify locally—previously, pushing to a test environment took 5 to 10 minutes, but now the full Manus runs on a local computer and verification takes only seconds; letting AI troubleshoot online issues by reading production logs, configurations, and exception records under permission controls; letting AI complete the release process, while approval authority remains with people; and letting AI perform end-to-end verification, using test accounts to go through real workflows and checking results in browsers and phone emulators.

The changes this system brought are very concrete. The Manus mobile team has always had only 7 people, yet it is responsible for 2 products and 4 mobile apps in total, with progress basically in sync with the web version. Product managers can also directly modify code. All product managers have submitted code to Manus, and about half can independently complete requirements and ship them. One person took a speech-to-text feature from model selection all the way to launch, and built both the web and desktop interfaces themselves. Video has become the new PRD: in a few minutes it explains the user scenario, interface design, post-click feedback, different states, and implementation logic.

The marketing team’s approach is closer to a digital colleague. After contract signing and internal approval are completed, the system sends a signal, AI updates progress, and finds the relevant colleagues to push the next step forward. This digital colleague judges whether a Slack message is relevant to it, proactively reminds people that a project is still in the testing phase, and asks across teams to clarify the applicable conditions for a new feature. When it is not working smoothly, it can also submit requirements through a feedback interface and automatically trigger a Manus development task.

In the comments, someone asked a practical question: who maintains the context? What if it expires and no one notices? This may be the most easily overlooked part of implementation. Writing a piece of context is not hard; the hard part is maintaining it continuously. Manus’s approach is to write processes and knowledge into Mono—the more people use it, the more accumulates. But keeping it fresh remains a long-term problem.

One commenter put it bluntly: the more detailed the instructions, the more the agent will only execute them literally. If the context is thick enough, it will fill in the level of detail itself. Another commenter mentioned that ByteDance has a management philosophy: increase talent density rather than set up complex rules and regulations. Applied to large models, that means using high-quality models plus more context plus fewer constraints.

Many teams still treat agents like interns, writing piles of detailed rules for fear that they will improvise. As a result, people still have to clean up the mess. Manus’s practice offers a counterexample: connect the materials, permissions, and processes properly, and let the agent run on its own. The human role has not become smaller; it has just become more concentrated. Define the problem, judge quality, decide what is worth doing, and take responsibility for the results.

That Netflix system of freedom and responsibility was originally written for employees. Now it seems agents may adapt to it earlier than people do.