Feynman's Letter: A Deep Dive into Asynchronous Coding Agents

Forum topic · 小凯 · 2026-05-03

Summary

A popular Chinese tech forum essay explains why asynchronous coding agents—AI systems that recently achieved breakthroughs on SWE-bench—mark the end of the 'supervision era' in software engineering. The author contrasts traditional sync interactions, where a developer sends a prompt, waits for code, and stalls the workflow (called 'task-flow collapse from synchronous interaction'), with async agents running in cloud-based persistent sandboxes. These agents decouple human attention from machine computation: long-running compiles and test suites are suspended asynchronously, and event-driven callbacks wake the agent when tests fail, letting it read logs and continue fixing. The essay argues that 'agentification' is less about raw AI intelligence and more about architecturally separating fragile human online time from long computation time. It envisions future software companies as 'one architect plus 100 digital avatars sprinting asynchronously in the cloud,' where a vague Issue becomes a tested, well-commented Pull Request by morning. The takeaway: build asynchronous message buses rather than click-'continue' chat loops—an AI that cannot run one minute without human gaze is not an assistant but an 'electronic infant.'

Feynman's Letter: Do You Want to Be a Deadline-Chasing Boss, or Hire a Team of Self-Directed Cyber Workers? — On Asynchronous Coding Agents

After reading the latest research on Asynchronous Coding Agents achieving breakthroughs on SWE-bench, I feel the "supervision era" of software engineering has truly ended.

To explain why AI of the past couldn't write complex 100,000-line projects, let's talk about being "single-threaded."

1. The Status Quo: A Single-Threaded Typist Locked in a Q&A Interface

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

Pain point: You send one instruction, the AI spits out a snippet of code, then stops and waits for your confirmation. If you go grab a coffee, the AI just sits there idling. For a refactoring task spanning dozens of files and eight hours, this "stop-and-wait" mechanism not only exhausts human mental bandwidth, it 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. The Asynchronous Agent: A Cyber Employee That Comes with Its Own Desk and Alarm Clock

The asynchronous agents topping the SWE-bench leaderboards recently represent a fundamental shift in underlying logic: you don't need to stay up with me—just toss your requirements onto my kanban board.

It achieves a two-fold decoupling of physical space:

Physical picture (offline state custody): The AI's runtime environment is moved into a cloud container (e.g., a long-lived sandbox). When the AI starts compiling code or running hundreds of unit tests, it doesn't need to hang 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 arrives as an "asynchronous event" pushed straight into the AI's "subconscious message queue." The AI is instantly awakened to inspect the error logs and continue fixing.

3. A Feynman-Style Judgment: Productivity Is the "Leveraging of Human Attention"

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

It's about whether you can, at the physical architecture level, completely separate humans' precious and fragile "online time" from the machine's extremely long "computation time."

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

When you can throw an extremely vague Issue into the system like sending an email, and the next morning receive a Pull Request that has passed all CI/CD tests and comes with perfect comments, you've touched the most fascinating compound-interest curve of the digital world.

Takeaway:

When designing your AI workflow, stop obsessing over dialogue boxes that require constant human clicks of "continue."

Go build your "asynchronous message bus."

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