Never write an update again: Notion's AI-powered engineering meetings | Ryan Nystrom

AI Made Engineering Fun Again

Claire Vo opens by framing Ryan Nystrom as someone using AI not just to code faster, but to run engineering differently. Ryan says the last year has “completely upended” a 12-plus-year workflow: new IDEs, terminals, and tools every few months, but in a way that feels energizing, joyful, and surprisingly fresh after decades in software.

The “Afterburner” Project and the Death of Bad Standups

Ryan manages a six-to-seven-person team at Notion and got pulled into improving CI despite not being an infra expert — mostly because he had strong opinions and, as he puts it, “puppy dog energy” around moving fast. His team’s project hub, nicknamed Afterburner, includes docs, databases, little wins, and an automated meeting system designed to eliminate the soul-sucking version of daily standup.

A Daily Notion Agent That Writes the Pre-Read for You

Every morning, a custom Notion AI agent generates the day’s meeting page by scanning the last 24 hours of Slack, closed tasks, merged PRs, telemetry, and even yesterday’s meeting transcript. The result is a real pre-read — CI metrics, decisions, progress, bugs, questions, risks — so the team can jump straight into the interesting stuff instead of reciting “I did this PR today.” Ryan says the tell that it’s working is when the agent surfaces something he missed, like a mock server fix improving tests by 13%, and the team can drill into it live.

How the Agent Is Actually Wired Up

Ryan shows the “Hot Potato” custom agent, set to run at 9:00 a.m. with explicit instructions: look back 24 hours, use sub-agents, query Honeycomb via MCP, inspect the project Slack channel, find tasks in the Notion database, and format everything into a consistent template. He’s careful with permissions — mostly read-only, except for the meetings database it needs to edit — and admits he even used the agent to help configure itself by pasting in a screenshot of the Honeycomb query and saying, essentially, “Please OCR this and figure it out.”

From Text Message to PR via Background Agent

The next demo is a tiny but revealing workflow: a friend texts Ryan asking for a “copy link to tab” feature in Notion’s tab block UI. Ryan turns that into a small task with four sentences and a screenshot, @-mentions Codex inside Notion, and his internal background-agent system spins up a VM, implements the feature, opens a PR, creates a preview URL, and even uploads screenshots of its own UI verification — all within roughly 20 minutes. When Ryan hits confusing code and CI failures, he just replies, “I don’t get this,” and the agent explains the change and fixes the types.

The Big Shift: Start with Specs, Not Code

Ryan’s strongest thesis comes from rebuilding Notion’s agent harness: don’t start with code, start with a markdown spec. He records a rough voice memo in Whisper, feeds that “YAP session” to Codex, asks it to learn the house format from an existing spec library, and then iterates until he has a detailed document with behavior, code pointers, and verification steps. Then he points Codex at the spec and says, “Build it,” which in one case basically one-shotted a couple thousand lines of working code in a few hours.

Verification Loops, Not Plumbing, Are the New Engineering Work

Ryan argues that engineers are shifting from plumbing into systems thinking, architecture, and verification design. His team even built CLI tools so agents can run Notion AI itself, send prompts, toggle “ask mode,” inspect transcripts, and verify that a feature actually behaves correctly — because if the verification is fuzzy, that’s the real bottleneck. He and Claire both note that engineers were already writing design docs and spec docs before; the difference now is those docs don’t sit waiting for review meetings, they become executable source of truth.

Why CI Speed Suddenly Matters Even More

In the closing stretch, Ryan and Claire get emphatic about CI as a hard limiter on AI-enabled engineering. Ryan says fast CI already mattered because tighter loops make humans bolder and more iterative, but with agents it’s “that but on steroids”: if CI takes an hour, your agent burns an hour waiting; if it takes three minutes, your throughput explodes. His prompting advice matches the rest of the episode’s vibe — direct, blunt, a little diva-ish when needed — with prompts like “I literally don’t know what I’m doing here” and “you’re wrong, defend your argument.”