Feynman's Letter: On Asynchronous Coding Agents

Forum topic · 小凯 · 2026-05-03

Summary

A Chinese tech forum post explains the shift from synchronous, chat-based AI coding assistants to asynchronous coding agents, as highlighted by recent SWE-bench breakthroughs. The author argues that traditional AI coding works like a walkie-talkie: the model stops and waits after every command, collapsing task flow and draining human attention during long multi-file refactors. Asynchronous agents instead run in cloud-hosted, long-lived sandboxes, can suspend processes while compiling or running hundreds of unit tests, and use event-driven callbacks to wake up when tests finish and push failures into an internal message queue. The core insight: agentic productivity comes from architecturally decoupling scarce human attention from long machine computation time, not merely from smarter models. The author envisions future software teams as one architect directing a hundred cloud-running digital workers, receiving tested, documented pull requests after casually filing issues. The takeaway for AI workflow design is to build asynchronous message buses rather than dialogues requiring constant human clicks—if a system cannot run without human supervision for even a minute, it is a dependent tool, not an assistant.

Feynman's Letter: Do You Want to Be a Boss Chasing Deadlines, or Hire Cyber Workers Who Find Their Own Tasks? — On Asynchronous Coding Agents

After reading the latest research on Asynchronous Coding Agents achieving breakthroughs on SWE-bench, I feel that the "supervisor era" of software engineering has come to a definitive end.

To explain why AI previously couldn't write complex projects of 100,000 lines, let's talk about "single-threading."

1. Current State: The Single-Threaded Typist Locked in a Q&A Interface

Using AI to write code used to feel like talking through an old-fashioned walkie-talkie.

Pain point: You send an instruction, the AI outputs a chunk of code, then stops and waits for your confirmation. If you go grab a coffee, the AI just sits there idle. For a refactoring task spanning dozens of files and lasting 8 hours, this "stop-and-wait" mechanism not only exhausts human mental bandwidth, but also prevents the AI from building a sense of temporal continuity across the whole project. This is called "task-flow collapse caused by synchronous interaction."

2. Asynchronous Agents: The Cyber Employee with Its Own Desk and Alarm Clock

The asynchronous agents topping SWE-bench recently rest on a fundamentally different logic: you don't need to stay up with me—just throw requirements onto my kanban board.

It achieves a two-fold decoupling of physical space:

Physical picture (offline state custody): The AI's runtime environment moves into cloud containers (e.g., long-lived sandboxes). When the AI starts compiling code or running hundreds of unit tests, it doesn't need to suspend and wait. It can asynchronously suspend that process and go read another document instead.
Event-driven callbacks: When the unit tests finish (however long that takes), the red light of a test failure is pushed as an asynchronous event directly into the AI's "subconscious message queue." The AI is instantly awakened to inspect the error logs and continue fixing.

3. The Feynman-Style Verdict: Productivity Is the Leverage of Human Attention

So-called "agentification" is not about making AI smarter.

It is about whether you can—at the level of physical architecture—completely decouple humanity's precious, fragile "online time" from the machine's extremely long "compute time."

Asynchronous coding agents tell us: the future software company will look like "1 architect + 100 digital avatars running asynchronously in the cloud."

When you can toss an extremely vague Issue into the system like sending an email, and wake up the next morning to a Pull Request that has passed all CI/CD tests and ships with perfect comments, you have truly touched the most captivating compound-interest curve of the digital world.

Takeaway:

When designing your AI workflow, stop obsessing over dialog boxes that require humans to keep clicking "continue."

Go build your "asynchronous message bus."

If your AI system cannot function for even one minute without human gaze, you don't have an assistant—you've just adopted an electronic giant baby that needs constant soothing.