Jesse Vincent: How I'm using coding agents in September, 2025 (blog.fsck.com)

Jesse Vincent lays out a practical, repeatable September 2025 workflow for using coding agents (primarily Claude Code) to design, plan, implement, and review features. He starts each task in an isolated git worktree (git worktree add …; npm install, lint, test) and keeps a CLAUDE.md with process rules. He splits agent work into roles: an "architect" session that brainstorms and writes a 200–300 word-at-a-time design, and an "implementer" session that executes bite-sized tasks from a detailed plan (docs/plans/) created with a planning prompt that assumes zero context, enforces TDD/DRY/YAGNI, and breaks work into tiny, testable commits. He alternates between implementer and architect—asking the architect to review completed tasks—resetting sessions to avoid context bloat. After implementation he pushes a PR that triggers CodeRabbit reviews and uses a coderabbit-review-helper to aggregate comments for the coding agent. This workflow matters because it shows a scalable, human-in-the-loop multi-agent pattern that reduces hallucination, context drift, and managerial overhead while preserving code quality. Key technical takeaways: isolate work with git worktrees, enforce incremental outputs and one-question prompts to keep agents focused, craft planning prompts that supply exhaustive developer context, and guard against agent credulity by role-playing review analysis (three-question prefix: hire, which issues to fix, are fixes correct). Risks remain—agents can accept bad reviewer suggestions—so Vincent emphasizes active PM oversight and cross-checks between architect and implementer.