Daily Brief: Once Coding Speeds Up, Management Becomes the Product Skill

Digg Tech’s strongest June 21 signal was Fiona Fung describing Anthropic engineers shipping far more code with agents, alongside a shift in hiring toward creative builders and deep systems thinkers. Lenny Rachitsky’s interview adds the management layer: routines, planning, and context-switch costs start to matter more when output expands faster than human attention. Anthropic’s recursive self-improvement essay provides the harder primary-source backbone here: more than 80% of merged code was authored by Claude as of May 2026, and the organization explicitly calls out human code review as a new bottleneck. The underlying pattern is that once execution gets cheaper, coordination becomes the expensive thing.

A lot of teams are still treating agent adoption like a productivity plugin for individual contributors. That misses the real change. When output rises sharply, quality, prioritization, architecture judgment, and team coherence become the gating factors. Product builders who learn to manage agents and humans as one system will outperform teams that simply generate more code, docs, or experiments without a stronger decision process.

Rebuild one team workflow around management, not generation. Trigger: a sprint plan, bug backlog, product spec, or cross-functional initiative. Context: business goal, acceptance criteria, technical constraints, roadmap priorities, and prior decisions. Tools: coding agents, docs, issue tracker, repo search, and a standing verifier such as tests, review checklists, or stakeholder signoff. Verifier: does the work match the intended goal, fit the system, and stay understandable by the next human? Budget: parallel-task limit, review capacity, token budget, and merge-risk threshold. Artifacts: plan, generated work, review evidence, and an explicit list of what was deferred. Stop condition: the work clears review without creating cognitive debt, or the loop pauses until scope and ownership are tighter.

For one week, cap the number of parallel agent-generated workstreams by your team’s actual review bandwidth. Then write down where the pressure shows up: specs, architecture, approvals, or context switching. That map will tell you whether your next investment should be in better prompts, better managers, or a better verifier.