Matt Pocock’s Agentic Engineering Workflow (just copy him)

Matt Pocock presents a framework where the primary value of AI in engineering is not the model itself, but the 'harness' around it. He argues that AI has consumed 'tactical programming'—the day-to-day writing of code—and that the remaining high-leverage work for humans is 'strategic programming': architecture, scoping, delegation, and workflow design. This is a significant shift for engineering leadership because it reframes AI adoption from a model-selection problem to an organizational design problem.

The core operational insight is that a well-factored codebase and clear task scoping allow cheaper, 'stupider' models to outperform expensive ones by reducing token waste on friction and ambiguity. This moves the cost conversation from API spend to architectural governance. Teams optimizing for token spend by switching models are optimizing the wrong variable; they should be optimizing for cohesion and modularity.

The video introduces a practical distinction between 'procedures' (human-invoked skills, like adversarial interviewing before coding) and 'abilities' (agent-invoked context). Pocock’s preference for explicitly invoked procedures signals a workflow design choice with major governance implications: keeping the human as the driver, not a passive reviewer, to prevent delegation of strategic thinking. His 'AFK agent' queue model—running sandboxed agents on GitHub Actions—is a concrete architecture for scaling parallel execution while maintaining review gatekeeping.

The teach skill demo, while compelling, is more a showcase of personal workflow and course marketing than a broadly generalizable framework. The advice to ignore model hype and wait a month before adopting new releases is pragmatically sound but anecdotal. Teams should watch for the tension between the 'bitter lesson' (trusting raw compute to beat human optimization) and the deliberate investment in human-centric harness design, as this will shape every AI tooling decision for the next 18 months.