Keynote
summary
Callstack replaces hand‑coding with agent orchestration using a CLI library called Agent Device, runs parallel agents via git worktrees and cloud agents, and repeatedly hits subscription token limits (weekly 20X Pro quota runs out by Tuesday). A specialized React Native model, Apex, follows a failed first run that hallucinated APIs, then uses curated data, a classifier to rank samples, and a multi‑stage pipeline to match state‑of‑the‑art on dedicated React Native Evolv tasks at a fraction of the cost, enabling ~100 PRs and 90% task coverage recently.
AI-written summary · about
chapters (18)
- 0:00Welcome to Agent Conf
- 0:50Mike Grabowski and Callstack's conference history
- 1:47Why Agent Conf exists
- 2:10How software development changed
- 3:50Michał Pierzchała takes the stage
- 4:11The model shift behind agentic coding
- 5:10Building Agent Device
- 6:33Guardrails for software quality
- 8:03Parallel agents and always-on workflows
- 9:08From agent workflows to a software factory
- 10:08Building independent AI infrastructure
- 11:40Why Callstack built a React Native model
- 13:14Curating training data after a failed first attempt
- 14:32Apex performance and training pipeline
- 16:03How Callstack uses Apex
- 18:26What unlimited tokens change
- 20:20The judgment engineers still need
- 21:07Closing and conference invitation
