Angie Jones on an Autonomous Engineering Organization and Its Consequences (Category AI4SDLC)

Yesterday I watched Angie Jones’s talk “Building an Autonomous Engineering Org.” It starts as a familiar AI-agent success story and ends with some very interesting questions about the future of such organizations. Jones is now VP of DevEx at the Agentic AI Foundation, but the talk draws on her time at Block. She says she spent the past couple of years helping transform an engineering organization of 3 500 people into an autonomous one. They succeeded, and at first it felt like a dream coming true… until it became a nightmare and ended with half the staff being cut as redundant under that degree of autonomy.

Returning to the first part of the talk, Angie argues that AI adoption by itself proves very little. Within a couple of months, around 90% of Block engineers regularly used Goose and Claude Code. There were metrics and token bills, but features were not reaching customers any faster. The company had moved through experimentation with adoption without reaching the point of obtaining value.

To describe progress toward an agentic engineering organization—where engineers use AI agents as their main means of achieving engineering outcomes—she introduces a scale of autonomy in the relationship between engineer and agent: — Stage 0: AI is not used in the workflow at all. — Stage 1: Autocomplete and similar suggestions, without agent mode. — Stage 2: Chatting with an agent, without real PRs. — Stage 3: The engineer delegates tasks to an agent and checks the results. — Stage 4: Several agents work in parallel. — Stage 5: An agent can take on an entire task and produce a production-ready result without constant human intervention.

They reached Stage 3 through AI champions rather than mass training for everyone. Jones selected around 50 engineers from critical teams who could spend roughly 30% of their time on AI enablement and knew how to work with nondeterministic models. Her point: a strategy that requires each of 3 500 engineers to become an advanced user independently will not produce a broad impact.

The champions first made repositories AI-ready. That sounds dull, but it is where autonomy begins: agents.md, claude.md, rule files, repeatable workflows, later agent skills, an AI code reviewer and attribution of AI tools in PRs. Agents need context, rules, build commands and an understanding of what counts as a good change in each repository. These changes were tailored rather than uniform: web, mobile, JVM backends, monorepos and small services needed different approaches.

Next, autonomy became native to the places where work already originates: Slack, Jira, Linear and GitHub issues. In one example, an engineer asks Goose about a bug directly in Slack. The agent visits the repository, confirms the problem and suggests fixes; the team chooses an option, and Goose returns with a PR. According to Jones, the cycle she describes—обсуждение -> диагностика -> согласование -> fix , or discussion, diagnosis, agreement and a fix—took around five minutes. This is a participant in the delivery loop, beyond an IDE coding assistant. Three months after the champions program began, AI-authored code had increased by 69%, reported time savings by 37%, and automated PRs by a factor of 21.

Stage 4 introduced a new bottleneck: when engineers produce 3-4 times as many PRs, review breaks first. Block made AI code review a mandatory part of the loop, using Codex on repositories and an auto-fix loop in which one agent identifies a problem while another fixes it and commits the change to the PR. They also needed cloud workspaces because engineers’ laptops could no longer handle several agents running in parallel.

Stage 5 required an organizational model, beyond another tool. The team began building Builder Bot: an orchestrator with a machine-readable model of the whole company, mapping services, their relationships and dependencies between codebases. Jones puts the number of repositories at roughly 25 000. Without that map, an agent can write a local patch but cannot plan a change spanning multiple products and systems. Technically, the story reached Stage 5 when any employee could ask Builder Bot in Slack to fix a bug or implement a feature, even without GitHub.

And this is where the talk ends with layoffs and questions such as: if we enabled people to build an autonomous engineering organization, could that become an argument for employing fewer people?

#AI #AI4SDLC #Engineering #Agents #Management #PlatformEngineering #Leadership #Processes