Ryan Nystrom, software engineer at Notion and co-founder of the acquired Campsite, is running Project Afterburner: a mission to cut Notion's CI time to one quarter of its current duration. The reason is not efficiency for its own sake. Slow CI is a hard ceiling on how fast AI coding agents can iterate, and that ceiling is now the bottleneck.
Nystrom's team uses an internal system called Boxy, a VM-based background agent that lets engineers @mention Codex directly from Notion comments. The result is a full pull request with screenshots delivered in 20 minutes. His spec-first workflow goes further: dictate an idea into Whisper, have Codex format it as a structured spec, commit that spec to the repo, and let the agent implement and verify it autonomously. The spec then becomes a versioned changelog for how a feature actually works.
The full conversation covers how to build a Notion AI custom agent that pulls from Slack, GitHub, Honeycomb metrics, and meeting transcripts to generate a daily standup pre-read, how to configure subagents and MCP integrations inside Notion AI, and how to prompt agents to defend their reasoning under pushback. Nystrom also makes a direct argument for why engineering managers and senior executives should keep writing code. The operational details here are specific enough to steal.
[READ ORIGINAL →]