Descriptions:
TypeScript educator and developer tools expert Matt Pocock sits down with David Ondrej to walk through his personal agentic engineering workflow and the philosophy behind it. The central argument draws on John Ousterhout’s “A Philosophy of Software Design” to distinguish tactical programming—the day-to-day writing of code—from strategic programming, which involves architecture, scoping, interface design, and velocity decisions. Pocock’s thesis is direct: AI has effectively consumed tactical programming, making strategic skills the primary lever separating developers who get enormous leverage from AI from those who see only modest gains.
On the tooling side, Pocock details his production setup using Sand Castle to run Claude Code (Opus 4) inside Docker or Podman containers, creating sandboxed environments that prevent agents from accessing sensitive local files or exfiltrating environment variables. This setup integrates with GitHub Actions to trigger automated agent review workflows on pull requests—agents check out branches, run type checks, and post review comments without any human intervention per cycle. The ability to parallelize agents in remote Vercel sandboxes, pulling commits back into a local workspace, is highlighted as particularly effective for high-throughput development.
Pocock pushes back on what he calls model obsession—the tendency to focus on which AI model is running rather than the quality of the surrounding harness. He argues that codebase architecture, test coverage, and documentation quality set the ceiling on what any agent can accomplish, making investment in those foundations more leveraged than chasing the latest model release. The conversation offers a practical, replicable framework for senior developers looking to get disproportionate output from agentic coding tools.
📺 Source: David Ondrej · Published June 18, 2026
🏷️ Format: Workflow Case Study
