Descriptions:
David Ondrej interviews Kun, a principal-level software engineer with background at Meta, Microsoft, and Atlassian, for an extended screen-share walkthrough of his complete agentic AI coding setup. The core of Kun’s workflow is a terminal-first environment built around Herder — a modern alternative to tmux — running on a Mac Mini accessible via SSH from any device. Rather than managing 20 or 30 parallel Claude Code sessions manually, Kun built a tool called FirstMate, which acts as an orchestrating agent that delegates to and monitors other agents, reducing the cognitive overhead of parallel agentic work to a single primary interface.
The conversation traces the evolution of AI-assisted coding from GitHub Copilot’s single-line completions through GPT-3.5 and GPT-4, landing on Claude Sonnet 3.5 v2 as the genuine inflection point where agents could take a full task and return complete results. Kun shares specifics about current model behavior: OpenAI’s “ultra” effort level triggers aggressive sub-agent fan-out where every spawned sub-agent itself runs at ultra, burning tokens rapidly, while “extra high” is his practical sweet spot for speed and cost. The pair also discuss a recent OpenAI billing error where the context compaction threshold for GPT 5.6 was misconfigured at 372K tokens, causing overcharges on any request exceeding 272K — OpenAI has since issued resets.
Kun flags DeepSuite as his most trusted current benchmark precisely because it is new enough to avoid training data contamination, and closes with pointed criticism of Anthropic’s communication style around Claude Code quota changes — noting that major policy shifts have repeatedly lacked enough clarity for developers to make informed decisions.
📺 Source: David Ondrej · Published July 17, 2026
🏷️ Format: Interview
