Kun Chen has observed a trend where some teams, particularly in smaller startups, are stopping manual peer reviews for pull requests to increase velocity.

According to Kun Chen, an engineer in an average software team typically writes 10 to 15 pull requests per month.

Kun Chen believes that current AI is not yet capable of replacing a smart human team for the purpose of ideation and brainstorming.

Kun Chen claims that using a fresh context window for an AI agent to review code is more effective at catching edge cases, as an agent reviewing in the same session it wrote the code is biased towards...

Kun Chen developed a tool called "No Mistakes" that automates the code validation and pull request process, including code review, testing, documentation updates, and risk assessment.

Kun Chen stated that the primary reason to use subagents is to prevent the context window of the main agent session from becoming too large, especially for isolated investigation or experimentation ta...

Kun Chen asserts that most current AI models are not proficient at proactively using subagents, with exceptions like Claude Code or Codex which may use a built-in "Explore" subagent for complex codeba...

Kun Chen built a tool called Treehouse to simplify the management of git worktrees by automatically setting up new workspaces from a managed pool where dependencies are pre-installed.

Kun Chen found that using HTML as an output format for AI agents is more effective than Markdown for human-agent collaboration because it can support richer, more interactive artifacts.

Kun Chen built a visual editor tool called Lavish that uses HTML to facilitate interactive planning and collaboration between a human and an AI agent.

Kun Chen stated that he generates 20 to 40 pull requests per day using his AI agent-assisted workflow.

Kun Chen stated that he typically has at least 5 different sessions actively running, with an average of 20 to 30 AI agents working in total.

The "test" phase of the No Mistakes tool functions like a CI pipeline, checking for regressions and generating evidence of functionality, such as screenshots or videos, for human review.

The ProgramBench benchmark, created by the developers of SWE-Bench, evaluates AI agents by tasking them with building complex programs like FFmpeg from scratch.

Claude Code has a slash command called "insights" that analyzes a user's sessions to generate a report with suggestions for improving their workflow and efficiency.