Descriptions:
Matthew Berman distills over 1,000 hours of hands-on experience with OpenAI Codex into a dense collection of advanced tips covering model selection, thread architecture, browser automation, skills, and long-running agent loops. The video opens with a counterintuitive finding about the GPT-5.6 model family — Soul, Terra, and Luna — showing on a cost-versus-performance graph that Luna Max outperforms Terra High while costing less, making Terra largely redundant. Berman’s rule of thumb: use Soul for the hardest problems, Luna for everything else, and avoid the “fast” mode unless time is critical (it costs 2.5x but delivers only 1.5x the speedup).
The bulk of the video covers Codex’s cross-thread delegation system, where individual threads can see all other threads in a project. Berman demonstrates spawning a deployment thread from a main coding thread using a GPT-5.6 Luna model — routing cheaper, well-scoped tasks away from the more expensive Soul model automatically via agents.md rules. A related pattern uses a supervisor thread to monitor a pool of parallel agent threads, restarting any that stall.
Additional sections cover Codex’s built-in browser (which Berman uses for DNS changes, Supabase autoscaling, and file management), skill installation including Matt Pocock’s coding skills library, and the goals/loops feature for autonomous long-running tasks. The video closes with advice on auditing agents.md after major model releases like GPT-5.6, since rules written for earlier models can actively degrade performance with newer ones.
📺 Source: Matthew Berman · Published July 14, 2026
🏷️ Format: Tutorial Demo
