Eric S. Raymond (ESR), the open-source figure who wrote The Cathedral and the Bazaar, posted on November 8 that he deliberately keeps his use of large language models "extremely simple, even primitive": no spawned subagents, no accumulated skills and no software factory that runs while he is away, just conversations in which he asks his "robot friend" to analyze and generate code. He says he has yet to see convincing evidence that this costs him productivity. His reasoning is that these elaborations cope with frontier models not yet being smart enough, and since models keep getting smarter, procedures with a short half-life are not worth the mental clutter. At writing time the post had about 1,300 likes, 127 replies and 26,000 views. In the same window, Box CEO Aaron Levie argued the opposite case for agent swarms.
Key Takeaways
- ✓Original post (@esrtweet, November 8): he just converses with the model to analyze and generate code, with no subagents, skill libraries or unattended software factory, and says he has "yet to see any convincing evidence" that this costs him productivity
- ✓Core argument: the elaborations cope with models not being smart enough yet, models keep improving, so the procedures have a short half-life; he calls this applying Rich Sutton's Bitter Lesson to his own workflow
- ✓Engagement (public X data at writing time): about 1,316 likes, 127 replies, 78 reposts, 214 bookmarks and roughly 26,000 views
- ✓The other side: Box CEO Aaron Levie posted the same day that agent swarms move productivity gains from "maybe 20-50%" to gains with no clear upper limit, across coding, cyber and life sciences
- ✓Middle ground: some developers replied that you should first understand the problem through conversation and only then build a factory, while others expect harness features to keep getting absorbed into the models
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the problem
AI coding tools in 2026 keep getting heavier: subagents, skill libraries, multi-agent orchestration and overnight software factories, and we have covered Coinbase CEO Brian Armstrong's chief-of-staff agent workflow. Many developers worry they are falling behind if they do not build a full system. On November 8, open-source figure Eric S. Raymond (ESR) posted the opposite answer, and it spread widely.
How it works
ESR's setup has almost no architecture: one conversation in which the model analyzes and generates code and he makes the calls. His argument has two steps. First, the elaborate procedures exist to cope with frontier models not yet being smart enough. Second, models keep getting smarter. If both hold, spending mental effort on procedures with a short half-life is a poor trade, and he would rather focus on the actual problems. He calls this applying Rich Sutton's 2019 Bitter Lesson, that general methods which scale with compute eventually beat hand-built cleverness, to his own workflow.
Benchmarks
This is an opinion debate. ESR gives no measurements, and "no productivity loss" is his own judgment. A useful reference is Garry Tan's open-source harness evals, which we reported earlier: seven setups on the same 20 tasks passed between 71% and 79%, and differed mainly in cost and speed. On the other side, Aaron Levie argues agent swarms push gains from "20-50%" to no clear upper limit, also without published measurements.
Getting started
Treat the debate as a self-check. Do the task in hand with the simplest chat workflow first, note which steps really keep getting stuck, and add subagents or automation only there instead of starting with a factory. Models improve noticeably every three to six months, so retest regularly and remove procedures the model now covers. For work that needs heavy parallelism, such as large-scale code review or security scanning, multiple agents still add clear value.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.