Ras Mic (Micky) walks Peter Yang through his 47-minute, human-in-the-loop approach to agentic engineering: automate the work, not the responsibility.

1. Isolate every task

  • Create a fresh Git worktree—a separate working directory and branch—for each feature or fix.
  • Fetch the latest main branch before starting; let several agents explore different features without editing the same files.
  • Worktrees do not isolate database state or development-server ports; handle those shared resources separately.

2. Build to a clear structure

  • Keep AGENTS.md focused on the workflow; let agents inspect the repository rather than repeating its entire layout.
  • Separate orchestration—business rules and when actions happen—from reusable service functions that do the work.
  • Give functions explicit inputs, structured results, and descriptive names to reduce duplicate helpers and make review easier.
  • This is Micky’s preferred architecture, not a universal requirement for every codebase.

3. Prove the change works

  • For a reported bug, reproduce the failure before fixing it; then capture the working result.
  • Use video for interactions, before-and-after screenshots for visible changes, and measurements or tests for nonvisual work.
  • His performance example shows navigation timings changing from roughly 850 ms to 60 ms; that is a demonstrated anecdote, not a general benchmark.
  • Evidence helps the agent notice weak changes and helps a human review faster; it does not establish complete correctness.

4. Ship through a bounded review loop

  • Open a pull request and run Greptile, an automated code reviewer.
  • Fix actionable feedback, push, and request another review until the target is 5/5 confidence with no unresolved comments.
  • The shared skill caps the loop at five iterations; his version permits ten, and pending reviews must remain visibly pending.
  • A reviewer score is not a safety certificate: Micky still tests features himself, and his consultancy also has developers review code.

Plan through conversation, preserve what must survive

  • Micky thinks aloud with the agent rather than routinely approving a formal specification.
  • An email-service discussion took about half an hour to clarify requirements such as tracking and custom domains.
  • For deferred work, he saves concrete instructions beside the code—an iMessage integration gets a setup.md file.
  • An early wrong assumption can spoil an entire overnight run; more autonomous execution does not fix unclear intent.

Models, tools, and operational limits

  • His model comparisons are personal experience: strong audits and computer use do not guarantee good taste or reliable skill execution.
  • Choose terminal or graphical tools according to the task; give interface work explicit visual references.
  • Schedule useful audits, but track jobs and clean up abandoned worktrees and dependencies.
  • Re-test new models without old instructions, then restore only the workflow rules they still need.
  • Use cheaper models where adequate; lab advice built around effectively unlimited tokens may not fit an ordinary budget.

“I’m not at a point where I trust the agent … especially with stuff that people are using.” — Ras Mic, 26:41