Ten lessons from three years building agents for real production work
garrytan · x · 2026-09-11
Garry Tan shares an article, "Ten Lessons Learned From Three Years of Building Agents," from a five-person team that has built agents for real production work since November 2023.
The team's architecture changed substantially over three years; some changes worked, others produced expensive and painful lessons. The article walks through those mistakes—an engineering retrospective from front-line production agent builders with direct value for anyone designing agent architectures.
More from coding & agent
- Early take: DeepSeek V4.1 Solo beats Agent Teams and GLM 5.3 Flash on quality and cost — teortaxesTex · 2026-09-11
- AI movie studio builds 3D camera previs tool to save generation credits — purecharisma2020 · 2026-09-11
- SpaceXAI team to livestream building a company from scratch with Grok next week — elonmusk · 2026-09-11
- Dev builds a 'Remote Debugger' with Astra to let Codex control another PC over LAN — ___Patrice___ · 2026-09-11
- Long-running subagent orchestrations still trip up most models, says dev — gabriel1 · 2026-09-11
- Wipe your Claude Code configs: dev says cleanup beat every new model release — tetsuoai · 2026-09-11