AI agents are just microservices with worse tooling

In this talk, Tyler Jang shares what Trunk learned while shipping an AI agent that fixes flaky tests. The thesis is that agents are not a new paradigm; they are distributed services, and platform engineers already know how to build those.

Key takeaways:

• Multi-agents beat mega-agents: smaller black boxes mean a smaller blast radius when models drift, with easier mocking and targeted evals.
• LLM-specific observability through classic instrumentation: daily trace review makes or breaks a team.
• Traditional SWE still wins: retries, structured contracts, and embedding clustering instead of prompts. The LLM should not do everything.
• What not to do: homebrewed audit logs, LLM-generated prompts, and letting the model run its iteration loop unchecked.

Attendees leave with patterns that work today, not hype.