Fragmented - AI Developer Podcast · Kaushik Gopal, Iury Souza

306 - Keeping your agent instructions in sync and effective

·23 min·1 clip

The paper found auto-generated agents.md files made results worse and used 20% more tokens.

1. Fragmented episode 306 is about keeping agent instructions in sync across AI coding tools and file formats. 2. Kaushik hosts the show with Yuri, and both talk from the perspective of Android developers moving into AI development. 3. The episode asks whether agents.md files are worth maintaining, and whether a single source of truth can reduce drift across tools. 4. The hosts describe the early mess of agent.md, clod.md, gemini.md, and hidden folders with extra instructions. 5. Kaushik says the industry is converging on agent.md, and connects that standard to the AAIF foundation. 6. They note that Anthropic remains a holdout with cloud.md, which keeps the ecosystem fragmented. 7. A web developer publicly asked why Anthropic could not consolidate around agents.md, and Tariq from Cloud Code replied with a non-answer response. 8. Yuri explains that the sync problem is not just project-level instructions, because he also wants global system-level instructions for his own style. 9. Kaushik describes using a separate repo for agent-related dotfiles, outside his original dotfiles repository. 10. Yuri says he uses rule sync, an open-source project that can import configurations from one harness and export them to another. 11. Kaushik says he previously kept a blog post updated and used a shell script to create symbolic links and set up folders. 12. The paper discussion starts with a claim that auto-generated agents.md files can make results worse and use 20% more tokens. 13. Kaushik adds that human-written files help by about 4% on average, but not consistently across models. 14. The hosts say the paper found agents do follow instructions, but unnecessary instructions create cognitive overhead. 15. They cite 10% to 22% reasoning-token consumption when a global agents.md conflicts with a local agents.md. 16. Kaushik notes that nested agents.md files can accumulate quickly because harnesses traverse directories and collect them. 17. The paper also argues that context files are often redundant with existing READMEs and developer documentation. 18. Yuri and Kaushik say the episode stays conversational and practical, with back-and-forth about what they actually use in their own workflows. 19. People managing Cloud Code, Codex, Cursor, or multiple agents.md files would get the most value from this episode. 20. People who want a purely theoretical AI overview without workflow details may skip it.

As heard by us

A practical look at keeping agent instructions aligned across tools.

Keeping agent instructions aligned is presented as a real maintenance problem, not a stunt, and Fragmented treats it that way. Kaushik and Yuri move through the global-versus-project split, dotfiles habits, a separate agents repo, and a sync workflow that cuts down on repeat…

Read the full review in PlayNext →

Why you'd press play

Stop rewriting the same agent rules in every tool.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Fragmented - AI Developer Podcast