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

I watched Angie Jones' performance yesterday.Building an Autonomous Engineering OrgIt begins as a common success story about AI agents, and ends with extremely interesting questions about the future of such organizations. Angie Jones now works for VP of DevEx at the Agentic AI Foundation, but in a report she shares her experience at Block, where she said she has been involved in transforming an engineering organization into a new and more advanced platform for the past couple of years. 3 500 Autonomous engineering organization. The guys did it, and at first it felt like a dream come true... until it became a nightmare and ended up cutting half the staff that had become redundant under such autonomy...

But if we go back to the first part of the report, Angie says that AI adoption alone proves almost nothing. Block in a couple of months about 90The percentage of engineers regularly used Goose and Claude Code, there were metrics and token bills, but features did not reach customers faster. The company went through the implementation phase, but did not reach the point of obtaining value.

And in order to talk about the degrees of movement to the agent engineering organization. (where engineers use AI agents as the primary way to obtain engineering outcomes) It introduces a scale of the engineer's autonomy relative to the agent. — Stage 0 AI is not used at all in the workflow. — Stage 1 Autocomplete and similar prompts, but without agent mode — Stage 2 Chat with an agent, but without real PR — Stage 3 The engineer delegates tasks to the agent and checks the result. — Stage 4 Multiple agents work in parallel. — Stage 5 The agent can be given a whole task and he is able to produce a result without constant human intervention.

To Stage. 3 They came through the use of AI champions, not mass training of everyone in a row. Jones chose roughly 50 Engineers from critical teams who could spend 30% of the time on AI enablement and were able to work with indeterminacy models. The idea is that if the strategy depends on each 3 500 The engineers themselves will become an advanced user, then a wide effect will not happen.

The champions first made AI-ready repositories. It sounds boring, but that's where autonomy begins: agents.md, claude.mdRule files, repetitive workflows, later agent skills, AI code reviewer, and attribution of AI tools to PR. The agent needs context, rules, build commands, and an understanding of what is considered a good change in that repository. All these changes were not the same for everyone, but took into account the specifics: web, mobile, JVM backend, monorepositories and small services required different approaches.

Then autonomy became native to places where work is already born: Slack, Jira, Linear, GitHub issues. In one example, an engineer directly at Slack asks Goose about the bug, the agent goes to the repository, confirms the problem, offers fix options, the team selects the option, and Goose returns with PR. According to Jones, the cycle обсуждение -> диагностика -> согласование -> fix It took about five minutes. It is no longer a coding assistant in an IDE, but a delivery loop member. Three months after the launch of the championship program, AI-authored code has grown in size. 69Percentage of reported time savings 37% and automatic PRs are in 21 once.

Stage 4 A new bottleneck: if the engineers start 3-4 PRs are bigger, review breaks first. Block connected AI code review as a mandatory part of the loop: Codex on repositories, an auto-fix loop where one agent finds a problem, another fixes and commits changes to PR. Plus, cloud workspaces were required: the engineers’ laptops ceased to withstand several parallel agents.

Stage 5 It was no longer a tool but a model of organization. The team began to build Builder Bot, an orchestrator with a machine-readable global model of the entire company: where the services lie, how they are connected, what dependencies between the codebases. According to Jones, it was about 25 000 repositories. Without such a card, an agent can write a local patch, but cannot schedule a change across multiple products and systems. Technically, Stage has ended. 5 Any employee could contact Builder Bot in Slack and ask to fix a bug or implement a feature, even without GitHub.

And here's where the report ends with layoffs and kind of questions: If we gave people the opportunity to build an autonomous engineering organization, could that be an argument for fewer people?

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