Descriptions:
Peter Yang interviews Kun Chen, a former Meta L8 engineer turned solo AI builder, who has developed a systematic method for shipping 20 to 40 pull requests per day using parallel AI agent sessions. Chen runs at least five concurrent Claude Code or Codex sessions at any given time, with each session potentially spawning multiple sub-agents, bringing the total active agent count to 20–30 simultaneously.
The conversation breaks Chen’s workflow into three phases: planning (human-led with AI assistance), coding (fully delegated to agents), and validation (mostly agents with selective human judgment). A central insight is that investment in the planning phase directly controls how long agents can run autonomously — detailed specs with measurable goals allow agents to iterate for extended periods without human re-prompting. Chen also explains his specific criteria for spawning sub-agents: primarily to prevent context window bloat in the main session by offloading exploratory or investigative work.
The interview addresses how existing team workflows — code review cadences, PR processes, sprint planning — were designed around human coding speed and break down when one engineer is opening 10 times the usual number of PRs. Chen argues that the bottleneck has shifted from writing code to managing agent orchestration and reviewing outputs at scale, offering concrete strategies for each stage.
📺 Source: Peter Yang · Published June 07, 2026
🏷️ Format: Interview
