The model is not the bottleneck: two essays on what still needs a human
Serguey Asael Shinder3 min readTwo essays published on 8 and 9 October come from different worlds — one from a software developer frustrated with coding agents, one from a mathematician writing to students who are abandoning PhD plans. Read together, they make the same argument from opposite sides: the model is not where the bottleneck is any more.
1. The agent around the model. Michael Lynch's "Why Are Coding Agents So Dumb?" separates the model (the LLM that writes code) from the agent (the software that connects it to files and commands, such as Claude Code or Codex). His claim: models kept getting better while agents "just stayed bad". His list is concrete — agents run embarrassingly parallel subtasks one at a time; never hand grunt work to a cheaper model or escalate hard work to a stronger one; do not know their own features; write "plans" that are piles of low-level decisions; stop for hours to ask what to name a git branch; and are only usable with access to everything, because their protections are polite instructions rather than OS-level sandboxes. His "table stakes for 2026" are deterministic filesystem and network boundaries, per-task model routing, and plans written for human comprehension.
2. The human around the model. Álvaro Lozano-Robledo's guest post on Terence Tao's blog answers students asking whether mathematics still has a future. His model is a "polytope of ideas": LLM proofs so far, however impressive, combine ideas already in the literature — they fill in the convex hull of what humans know. He expects rapid gains as the hull is filled and then a slowdown, because extending it needs a new definition or concept, which so far has come from people. He adds a practical observation: his own research projects tripled in a few months, and he is recruiting more students than ever, because LLMs widen what a researcher can attempt. Writing after OpenAI's large mathematics release, he notes the company reported attacking about 4,000 open problems and making progress on about 700 — remarkable, and also a measure of the limit.
The common point. Both essays reject the framing that capability lives in the model alone. Lynch locates the missing value in orchestration and safety — engineering that is decades old (schedulers, sandboxes, cost-aware routing) and simply not built yet. Lozano-Robledo locates it in new ideas and in people who can digest and judge results. Neither is a claim that models are weak.
What to do with this. For engineering teams the takeaway is testable: measure where your agent time is lost — waiting, serial execution, re-asking — rather than switching models; and treat any agent without an OS-level sandbox as having access to everything it can reach. For anyone advising students, the second essay's answer is the useful one: the reason to learn a field deeply has not changed, and the people who can extend the hull, or check what was found inside it, are the ones the tools make more productive.
This analysis draws on: Michael Lynch, "Why Are Coding Agents So Dumb?", 9 October 2026 — https://mtlynch.io/why-are-coding-agents-so-dumb ; Álvaro Lozano-Robledo, "What should we tell our students?" (guest post on Terence Tao's blog), 8 October 2026 — https://terrytao.wordpress.com/2026/10/08/what-should-we-tell-our-students