L8 Principal's Agentic Engineering Setup (just copy him)

David Ondrej interviews Kun about a solo-development workflow with one coordinating agent and many parallel sessions. It moves routine coordination out of the human loop while returning ambiguous decisions to Kun and validating consequential changes before merge.

Key Points Covered

• Use a coordinator instead of personally juggling sessions: Kun's First Mate receives his requests and routes work to crew agents in Herder. It lets him inspect an individual session only when direct intervention is useful [00:00:00]-[00:01:03], [00:08:53]-[00:13:52].
• Reserve human attention for ambiguity: Kun describes his day as dumping thoughts into First Mate and handling decisions it cannot make. He says explicit escalation instructions keep autonomy from silently drifting in the wrong direction [00:21:01]-[00:22:47].
• Turn complex design reviews into interactive artifacts: For a workspace-organization decision, Lavish produces an HTML review and editable diagram. Kun can inspect alternatives and send a selected trade-off back to the agent [00:13:52]-[00:15:47], [00:18:53]-[00:21:01].
• Route models according to task and quota: First Mate follows rules for selecting models and reasoning levels. Kun tracks subscription limits and argues that background jobs need a cheaper slow mode rather than premium low latency [00:15:47]-[00:18:53].
• Validate changes through a separate pipeline: No Mistakes derives requirements from the original agent session, rebases, and performs adversarial review. It fixes clear defects or escalates product implications, then tests, checks documentation, and watches CI [00:37:41]-[00:41:34].
• Spend review effort according to risk: Kun treats validation as a cost of quality for production systems. He skips the heaviest process for some demos or simple changes. His dashboard attributes many caught issues to adversarial review and stale documentation [00:42:33]-[00:45:28].
• Build workflow tooling from observed bottlenecks: He says First Mate stopped setup work from consuming his product time. He says the broader toolchain emerged from concrete problems in coordinating sessions, reviewing generated code, and planning visually [00:45:28]-[00:49:36].
• Optimize interfaces for agents, not just humans: Kun expects more services to become headless. He presents benchmarks where CLI-shaped AXI wrappers reduce turns, latency, or token cost compared with the MCP interfaces he tested [00:49:36]-[00:56:47].

Full video: https://www.youtube.com/watch?v=8ZgpAXe5V5w(opens in a new tab)