SpaceX engineer describes always-on agents that could run for weeks

Lee Robinson, a SpaceX technical lead and former Cursor core member, delivered a lecture at Stanford University's CS146S course on the working principles of always-on agents, sharing his views on the next phase of AI agent development.

Robinson argued that every new product must include some form of agent capability, either integrated into existing systems or it will become headless data for other agents. He said the ability to serve as an interaction hub for agents will increasingly determine a product's value.

He divided developer-AI collaboration into four eras: copying and pasting ChatGPT code, interacting with models in the terminal, building dedicated coding agents, and now an always-on era in which agents have their own computer and memory. In this model, an agent acts more like an always-available colleague, running in the background for hours or days without requiring step-by-step instructions.

On cost control, Robinson described principles including thin-client fixed-server architecture, main-agent/sub-agent coordination, and prompt caching. He said a main agent can delegate heavy tasks to sub-agents with fresh working memory to work around context-window limits, while teams should keep prompts cached and compress long conversations before cache expiry to reduce token costs. He also shared the 'delete the product' principle used at Cursor and SpaceX: assume today's models are the worst they will ever be and build minimal interfaces that let models think and act.

Robinson predicted that within the next six to twelve months, improvements in long-horizon instruction following and persistent workflow tools will allow agents to work continuously for weeks or months, posing new challenges for model evaluation.