David Ondrej

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

⏱ 62 min video · 3 min read17 Jul 2026Worth watching

TL;DR

A senior engineer (formerly at Meta, Microsoft, Atlassian) walks through his complete solo AI development setup built around a custom orchestrator agent called First Mate, which manages dozens of parallel Claude/GPT sub-agents so he only deals with genuinely ambiguous decisions. He also demos No Mistakes, his adversarial AI code-review pipeline that caught mistakes in 63% of 1,000 changes across 59 repos.

Key points

First Mate is a custom orchestrator agent (agents.md + bash scripts) that coordinates all sub-agents, letting him interact with just one agent while it delegates tasks to crew mates using Claude Code or pi with GPT-4.1/5.6

No Mistakes is his adversarial review pipeline: it analyzes original intent from the agent session, runs GPT-5.6 Soul as an adversarial reviewer, auto-fixes obvious bugs, escalates product-impacting bugs, then does testing, linting, docs, and babysits CI until PR is clean

He uses Zellij (Herder) as his terminal multiplexer instead of tmux because it natively understands agent sessions and shows working status, and Wezterm as his frameless, fully customizable terminal

His model routing rules send high-complexity design work to Claude Opus (Fable) in Claude Code, default daily work to GPT-5.6 Soul on extra-high reasoning in pi, and speed-critical tasks like home automation to GPT-4.1 Luna

He argues the $200/month subscription tier is insufficient for serious agentic work and advocates for a ~$500 slow-but-cheap tier for background tasks, noting API pricing would cost him over $10,000/month

Actionable insights

→

Use a single orchestrator agent (like First Mate) as your only interface and let it delegate to sub-agents — this eliminates tab-juggling and keeps you in a brain-dump mode focused only on decisions that need human judgment

→

Run an adversarial review pipeline (different model reviews AI-generated code) before merging — GPT-5.6 Soul on medium was the most effective reviewer, catching edge cases and stale documentation, and flagging 63% of 1,000 real changes

→

Set explicit rules for when your agent should escalate vs. act autonomously — the sweet spot differs per person, but getting this calibrated prevents agents going rogue while still minimizing your involvement

→

Use Lavish (an interactive HTML artifact API) to turn complex technical design discussions into visual whiteboards you can annotate and send feedback back to the agent, instead of reading walls of terminal text

→

Use extra-high reasoning on GPT-5.6 Soul rather than ultra — ultra triggers aggressive sub-agent fan-out that burns tokens fast, while extra-high is both faster and more economical for most tasks

Notable quotes

“I don't want to spend all day just juggling between the tabs right and remembering what was what. I developed First Mate because I think the agents as they become more capable should be able to juggle all those things for me.”

“If I actually take API pricing for everything I worked on for the past month, it is going to cost more than like $10,000. It is just not sustainable.”

“Now this software is pretty much like unstoppable. There is no way you can stop First Mate from doing what it needs to do. Even if the scripts have various kind of bugs, it will make First Mate a little bit more inefficient, it will not stop it from working.”

Worth watching?

✅

Worth watching the full video?

Watch if you want to see a working, battle-tested multi-agent engineering system built from real solo-founder pain points — the screen-share walkthroughs of First Mate and No Mistakes are uniquely detailed and hard to find explained this concretely elsewhere.

Topics

AI & Tech Claude Code

Explore more summaries on these topics →

Saved you some time? The creator still deserves a like.

Watch on YouTube →

More like this