Steve Yegge

Redefining the Engineering Workflow Simply purchasing access to frontier AI models does not make a company ship features faster. When Angie Jones took on the task of building an agentic organization at Block, she discovered that 90% of her engineers were already using AI inside their IDEs. Yet, product delivery speeds remained entirely unchanged. Real impact requires moving past simple code generation toward true workflow integration. To bridge this gap, she defined an agentic engineering organization as one where developers do not just write code with AI, but actively direct agents. This operational shift forces engineers to act as managers: decomposing complex problems, delegating tasks, and rigorously reviewing machine-generated output. The Six-Stage Maturity Model To map this transition, the organization adapted a maturity framework inspired by Steve Yegge's observations on "Gastown." This model tracks the shifting relationship between human and machine across six distinct levels: * **Stage 0:** Complete manual coding without AI assistance. * **Stage 1:** Basic auto-complete operations inside the IDE. * **Stage 2:** Chatting with agents without generating pull requests (PRs). * **Stage 3:** Delegating specific tasks to agents and checking the output. * **Stage 4:** Running multiple specialized agents in parallel. * **Stage 5:** Full task delegation where agents produce shippable code autonomously. Most developers naturally stall between stages one and two. Bridging the gap to stage five requires systemic, structural changes to the codebase rather than telling individual developers to work harder. Repository Readiness and the 1% Strategy Instead of forcing 3,500 engineers through a top-down mandate, the initiative focused on a handpicked group of 50 power users representing critical repositories. These "AI champions" spent 30% of their time making codebases AI-ready. They embedded context files like `agents.md` or `claude.md` alongside strict rule files to act as guardrails. This customized approach accommodated diverse repository shapes, ranging from massive Java Virtual Machine (JVM) mono-repos at Square and Cash App to nimbler mobile setups at Tidal. Eliminating Bottlenecks in Parallel Production When agents began working directly inside Slack, Jira, and Linear, PR production skyrocketed. However, this sudden surge created massive bottlenecks. Code reviews stalled, and local laptops choked under the processing load. To stop the bleeding, the team deployed Codex to automate initial code reviews and implemented automated self-healing fix loops. They also transitioned operations to isolated, cloud-based workspaces to let multiple agents run in parallel without crashing local systems. Ultimately, the team built an internal orchestrator called Builder Bot, powered by a 25,000-repository global world map, which allowed any employee to deploy features directly from chat.