GPT-5-Codex-Max: Training Agents with Personality, Tools & Trust
Brian Fioca and Bill Chen from OpenAI join Latent Space at the EIE Code conference to discuss the newly launched GPT-5-Codex-Max, a coding agent designed to run for 24+ hours autonomously. They explain the naming choice ("Max" for maximalist speed/performance vs. "Pro" for slower, more thoughtful), and how the model was trained with a specific "personality"—defined by behavioral characteristics like clear communication, planning, context gathering, and checking work—that developers can trust as a pair programmer. A central theme is the distinction between the Codex line (a frontier coding model optimized for its own open-source harness with firm opinions on tools) and the mainline GPT-5 models (more general, steerable across diverse tool environments). They share concrete observations about how models develop "habits"—for example, Codex performs noticeably better with tools named "rg" (ripgrep) rather than "grep"—and discuss the broader trend of the abstraction layer moving up from models to agents, enabling sub-agent patterns where agents spawn other agents (e.g., a Codex instance that writes custom integration plugins). The conversation also covers OpenAI's applied evals initiative, including challenges around multi-turn evals, the use of LLM-as-judge over full trajectories, and a proposed "job interview eval" framework for evaluating agents on underspecified, real-world tasks. They close with 2026 predictions: deeper computer use (especially for legacy apps with only UIs), more capable vision, higher trust levels allowing smaller teams to access top-tier engineering capabilities, and continued packaging of complete agents rather than raw models for developers to integrate.
Original sources
Podcast Latent Space
YouTube Latent Space
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.