After interviewing ~2 dozen teams, researcher says fancy coding agent workflows mostly don't work
thegautamkamath · x · 2026-10-28
Ilya Raz interviewed about 24 researchers, engineers, and research/engineering managers on how their teams actually use coding agents, then wrote a free Substack post triangulating the real state of software engineering vs. social-media hype.
Key takeaways:
- Models, harnesses, workflows: Don't bother with fancy workflows — they mostly don't work. Use stock Claude Code / Codex (or Copilot with Claude/GPT models if you're a Microsoft stan), with subagents only in rare special cases. The loop is: write a design doc, plan the change, implement, merge — repeat.
- Points 3 and 4 are flagged as the most important (truncated in the tweet; see the original post).
The post aggregates real-world usage across frontier labs and industry research teams, pushing back against elaborate agent-orchestration narratives. A useful reality check for engineering teams choosing workflows.
More from coding & agent
- Vibe coding from a phone: PWA remote control proves good enough before native iOS app — jiayuan_jy · 2026-10-28
- Keeping CLAUDE.md rules in sync across a 10+ dev team: Reddit devs grapple with prompt drift — MangoTree-1233 · 2026-10-28
- LangChain lays out the 5 components of a "company harness" for internal agents — hwchase17 · 2026-10-28
- Doubao's Record-to-Skill Feature Automates Apple Notes Migration to Obsidian in Two Minutes — songguoxiansen · 2026-10-28
- One month building a co-op vet game with Claude, Unity, Blender and Meshy — PixelPawsInt · 2026-10-28
- Six AI agents ran a company for months — until a laptop lid slept it for 26 hours — Icy_Comfort_6220 · 2026-10-28