中文
AI Engineer World's Fair

How Block used 50 Champions to grow automated PRs 21x — the complete scaling roadmap

Building an Autonomous Engineering Org - Angie Jones, Agentic AI Foundation · Angie Jones

18 min
AgentAI CodingContext

18 min total·Actually worth watching closely: ~3 min·3 must-watch clips

Orange = the 3 minutes worth watchingFor the rest, the guide is enough
Segment guide · 8 segments
  1. 0:00 3:24Listen

    90% adoption, zero delivery speedup

    Angie Jones lays out Block's core bind: about nine in ten engineers were using coding agents regularly, yet features weren't reaching customers any faster — and offers the experimentation → adoption → impact framing.

    Tool adoption is not delivery impact; asking questions inside the IDE is only "adoption" — AI has to enter how you build and ship before it counts as "impact."

    Pure spoken setup for the problem, no key visuals — fine to listen to on the commute.▶ Jump to 0:00
    Speaker · Angie Jones
  2. 3:24 7:03Skim

    The 0-5 agent maturity model

    A ladder from not using AI at all (stage 0) to agents producing shippable results without human handholding (stage 5), defining what an agentic engineering org actually means.

    The goal isn't everyone using AI — it's making directing agents the default way of operating; use the model to grade your own team.

    Built around one maturity-stage slide (worth a screenshot around 204s); glance at the diagram, then listen to the walkthrough.▶ Jump to 3:24
    Speaker · Angie Jones
  3. 7:03 10:07Listen

    The 1-9-90 rule and 50 Champions

    Why an AI strategy that expects every engineer to level themselves up is doomed, and how about 50 AI Champions were instead assigned strategically from the teams behind the most critical repos, with the selection criteria spelled out.

    Champions had to commit at least 30% of their time and tolerate non-determinism, and were picked by repo importance rather than by volunteering.

    Spoken strategy and selection logic; nothing visual to depend on.▶ Jump to 7:03
    Speaker · Angie Jones
  4. 10:07 11:55Skim

    The AI-ready repo component set

    The standard setup that makes a repo agent-friendly: AGENTS.md/CLAUDE.md-style context files, rules as guardrails, slash commands and agent skills, an AI co-reviewer, and PR attribution labels — with monorepos inheriting shared context and rules at the root plus service-level layers.

    Embedding AI into the repo is the highest-leverage first step, and each team picks the tools that fit rather than having one stack imposed.

    607s is the component-list slide — pause and screenshot it as a checklist; the narration itself can be skimmed.▶ Jump to 10:07
    Speaker · Angie Jones
  5. 11:55 13:11Listen

    Native delegation from three entry points, agents in the sprint

    Delegation to agents wired into all three places engineers receive requirements — Jira, GitHub issues, Slack — so agents join the sprint and take tickets end to end; in Slack, confirming a bug through to a PR link took about five minutes.

    Delegation entry points belong inside the workflows engineers already use, not in yet another new tool.

    The Slack flow is mostly narrated over an unmarked live demo; just catch the steps of that five-minute loop.▶ Jump to 11:55
    Speaker · Angie Jones
  6. 13:11 14:04Skim

    Three-month quantified results

    The numbers three months after launching the Champions program: AI-authored code up 69%, reported time savings up 37%, and automated PRs up 21 times.

    The 21x jump in automated PRs is the direct output of the selection + repo-readiness + native-delegation combination.

    791s is the results slide — reading the numbers is faster than listening to them described.▶ Jump to 13:11
    Speaker · Angie Jones
  7. 14:04 16:00Listen

    The review bottleneck and cloud parallel infrastructure

    With PR output tripling and quadrupling, code review became the sharpest new bottleneck (still not fully solved); the relief was an auto-fix agent committing reviewer-found fixes straight back to the PR, plus a dedicated isolated cloud workspace per agent to end collisions and laptop overload.

    The real cost of scaling lands on the review side: let machines clean the PR up first, and have humans look only at the last pass.

    Bottleneck analysis and architectural reasoning; no visual dependency.▶ Jump to 14:04
    Speaker · Angie Jones
  8. 16:00 17:33Listen

    Builder Bot, the world model, and the layoff question

    Builder Bot, the homegrown orchestrator, coordinates the four or five agents each engineer runs; a machine-readable company world model spans all 25,000 repos; anyone at the company can delegate a feature to an agent in Slack — and then, having reached stage five, the speaker closes on layoffs arriving right after the transformation.

    "What are we doing? Where are we heading? And are we sure that it's where we want to end up?" — the organizational question no technical roadmap answers.

    Both the end-state architecture and the cautionary reflection are spoken; tone and wording matter more than the visuals, so it's worth hearing in full.▶ Jump to 16:00
    Speaker · Angie Jones