SpaceX Engineer Tan: Agents Need Layering
SpaceX AI engineer Lauren Tan (social account @poteto) recently held a Q&A (AMA) session, sharing her personal practices in using and managing an AI Agent team.
The AMA primarily featured Tan herself answering questions directly, with a brief appearance of a six-person video conference, without the use of a whiteboard for explanations or providing a recording playback.
Tan introduced that multiple bots can share the same underlying computer, but their memory data and chat records are independent and isolated from each other, preventing any mixing.
For complex tasks, she suggested adopting a "specialized team" model, with a "Chief of Staff" role as the sole task entry point, responsible for receiving requests and assigning them to the corresponding specialized bots; her own bot takes on an "outer loop" role, responsible for integrating with external systems like Slack, email, and calendars, and then triggering the "inner loop" running in Cursor to further activate cloud agents to execute specific tasks.
Tan mentioned that currently, the outer loop bot cannot freely switch underlying models, but the cloud agents it dispatches can choose any model; she illustrated the overall workflow: monitoring the Slack feedback channel triggers the inner loop in Cursor, which then starts the cloud agents in "potato mode" to execute tasks.
Tan summarized her Agent team-building philosophy as "first cultivate advisory agents, then expand to execution worker agents," and reminded users to avoid becoming a "task chaser" who constantly urges various bots to execute tasks throughout the process.
This practice reflects the current evolution of AI Agent engineering from "single chatbots" to "multi-agent systems with role division and scheduling hierarchy"—an agent architecture with a unified entry point, task assignment, and inner and outer loop scheduling capabilities can significantly reduce the user's task coordination burden. The dissemination of such methodologies will also encourage more companies and developers to adopt similar "advisory first, workers later" paths in their workflows, objectively benefiting the development tools and platforms that provide multi-agent orchestration and task scheduling infrastructure.
Source: Public Information
ABAB AI Insight
As a company known for its engineering culture of high-intensity automation and toolchain efficiency, SpaceX's internal AI team has been continuously integrating self-developed and third-party AI tools (such as Cursor, an AI-native code editor) into daily R&D processes. Tan's sharing of multi-bot collaboration and layered scheduling methods continues SpaceX's consistent engineering culture path of "replacing repetitive human coordination work with automated tools."
From the perspective of industry resource investment, AI-native development tools like Cursor and Claude Code from Anthropic have been continuously increasing their R&D investment in "cloud agents" and multi-model scheduling capabilities. Their core motivation is to shift from a single conversational AI assistant to an agent infrastructure capable of autonomously executing complex task chains, aiming to capture market share in enterprise-level AI development tools. The "inner loop triggering cloud agents" workflow described by Tan is built upon the foundational capabilities that these tool vendors continuously invest in.
This is similar to the earlier evolution in software engineering from "monolithic applications" to "microservices architecture"—at that time, companies split a single large system into clearly defined, independently scheduled service units to enhance maintainability. The current AI Agent engineering practice's design idea of "specialized teams + Chief of Staff as a unified entry point" essentially migrates the same architectural thinking from software systems to the organizational method of AI Agent teams.
Essentially, this represents a technological substitution— as cloud agents can freely call any underlying model and can independently complete the entire chain from task reception to execution, the user's role is shifting from "an operator urging bots to execute tasks one by one" to "an architect designing scheduling architecture and defining responsibility boundaries." The driving mechanism is that as the number of agents increases and task complexity rises, the marginal cost of human coordination will quickly exceed the fixed cost of designing an automated scheduling architecture, forcing practitioners to shift towards higher-level architectural design work.
ABAB News · Cognitive Laws
As the number of agents increases, you should transition from operator to architect.
Separate memories allow multiple bots to share a single brain.
First nurture advisors, then expand workers, and avoid becoming a task chaser yourself.