⚡️GPT5-Codex-Max: Training Agents with Personality, Tools & Trust — Brian Fioca + Bill Chen, OpenAI
Summary
- Codex Max is pitched as both long-running and faster: it can work for “24 hours or more,” manage its own context through compaction inside the Codex harness, and sometimes reach the right answer faster on the same class of problem. Brian has run it locally for longer than a day across several days, with his laptop closed and nobody opening it; “Max” means “speed and maximization, like maximalist,” not merely slower deliberation.
- OpenAI’s product direction increasingly treats performance as a model–harness package rather than an interchangeable raw model. Codex is optimized for terminal-shaped tools, while mainline GPT-5 is broader and more steerable; Brian says Codex is open source and its model is available in the API. Partners found that naming a search tool “rg” rather than “grep” made tool-call performance “jump by a lot.” The host says generalization remains the end goal, framing GPT-5’s generality and Codex’s coding focus as two horizons.
- “Personality” here means reliable engineering conduct—planning, context gathering, progress updates, and checking work—because trust constrains delegation. GPT-5.1 can preface tool calls and be prompted to suppress chatter, while Codex relies on a reasoning summarizer. Communication tokens may be wasteful in a 24-hour cloud run, but they help engineers follow, interrupt, or stop bad trajectories early.
- The abstraction layer is moving from models to packaged agents that other products can invoke, reducing the need to retune a harness for every release. Zed, GitHub, and VS Code are cited as surfaces that can use a whole agent. Codex Max’s context management also supports agents handing work to or spawning parallel subagents, although the pattern is still emerging.
- Applied evals are the operating system for trust, product improvement, and model training—not a leaderboard sidecar. OpenAI wants customer-specific failures and evals, using traces, graders, guardrails, and metaprompting to improve behavior; Bill’s analogy is a “PhD in an API” who still needs a job description, mentorship, and performance reviews. Brian says initial Codex adoption at OpenAI was around 50%; Bill says those users use it every day.
- Coding agents are breaking out of coding into general computer automation, although vision and application access remain bottlenecks. The guests frame a coding agent as “a computer use agent but for the terminal,” already useful for email, media snippets, experiment directories, and desktops. Bill names Devin and Cascade as targets to beat and describes building a Devin for non-coding/email; their 2026 hopes include more computer use, broader subagent systems, UI-only integrations, and enough trust to give ordinary teams capabilities associated with top-tier firms.
Deep dive
1. Codex Max combines a long horizon with speed
Brian’s explanation of “Max”: Pro could imply slower, more thoughtful work, whereas Max signals “speed and maximization, like maximalist.” It can run “24 hours or more,” yet for the same problems it can also reach the correct answer faster.
His strongest endurance specimen was local: a run lasting beyond 24 hours over several days, with his laptop closed and nobody opening it. Inside the Codex harness, Max compacts and manages its own context, letting it “run basically forever” without manual context-window management.
2. Trust is trained as behavior, while tools expose model habits
Brian says he was close to the training team while GPT-5 was training. His framing of personality for a trusted pair programmer is concrete: communicate what it is doing, plan when appropriate, gather context before diving in, and check its work. These are “best software-engineering practices” converted into measurable behavioral characteristics.
The host’s pushback is whether personality matters for an unattended “cron job” lasting 24 hours. Bill answers that engineers currently want progress updates so they can interject or stop the agent, or at least avoid wasting time on a rollout they will have to throw away.
For GPT-5.1, OpenAI uses preambles such as “I’m about to go look for this,” and users can steer or suppress them. Bill says he gives his personal agent a fun “buddy” personality, while acknowledging that such chatter burns unnecessary tokens in long cloud jobs; Codex instead relies on its reasoning summarizer for updates.
Bill’s sharpest example of harness dependence: Codex was trained around terminal tools, but partners retained other tools by matching terminal-style names and input/output shapes. A search tool called “rg” performs better than one called “grep” because “Codex loves ripgrep.” The host objects that models should generalize and calls that the end goal, while framing general GPT-5 and coding-focused Codex as two horizons.
3. The product boundary is shifting from models to agents
Brian describes “the abstraction layer really moving upwards” toward the agent layer. Instead of rebuilding around every model and API release, developers can embed a packaged Codex agent whose model, tools, sandboxing, and harness have been designed together. Brian also says Codex is open source and its model is available in the API.
Bill names Zed, GitHub, and VS Code as using this pattern. A coding-tool builder can avoid maintaining a team around every model or API release and every harness, sandboxing, or tool change by building one layer above the packaged agent.
Codex Max’s context management also supports agents using agents: it can hand context to subagents, spawn parallel work, and create new abstractions during a long-running workflow. Both guests stress that the primitives are only now being built, so the ultimate operating pattern remains uncertain.
4. Evals turn autonomy into an inspectable production system
The trust claim is already behavioral, not hypothetical. Bill says he has not written a line of code by hand in months and shipped an open-source Codex upgrade pack for migrating from Completions to Responses without manually writing its code. Brian reports that around 50% of OpenAI folks initially started using Codex; Bill says they use it every day.
Bill argues that academic benchmarks leave a gap around “what people care about the most.” Applied evals capture customer use cases where one missing capability blocks deployment, giving OpenAI concrete deficiencies to “hill-climb on together” through model and product changes.
Bill’s analogy is that calling a model a “PhD in an API” is incomplete because a new hire still needs a job description—the prompt—plus mentorship, guardrails, and performance reviews. Agent traces, rollout traces, graders, and production checks make it possible to identify a behavior, have the agent improve its own instructions, and test the next run.
Multi-turn evaluation remains unsettled. Bill suggests judging an entire trajectory, walking back to a weak step, and rerunning with improved instructions; his proposed “job interview eval” would reward an agent for clarifying an underspecified task before implementation. The host’s concrete request is batch multi-turn evals, because thousands of non-time-sensitive runs should execute overnight as cheaply as possible; the discussion says that capability is not yet available.
5. Coding agents are becoming terminal-native computer users
When asked what they are trying to beat, Bill names Devin and Cascade and says he is building a Devin for non-coding work, especially email. Brian calls Slack “the ultimate user interface” for work; Bill says he interacts with his email agent through Slack.
The guests’ broader framing is that coding tools are becoming personal automation. Codex can help sort email through terminal utilities, generate video snippets, organize experiment directories, or clean a desktop. Bill connects this to his 1990s systems-administration work, when Bash scripts and custom software solved real-world tasks beyond writing code.
Bill calls coding agents “a computer-use agent but for the terminal,” but says they are not yet vision-native enough. Many legacy or closed applications expose only a UI rather than an API or MCP, making computer use important for accessing user-owned data.
Their 2026 hopes include more computer use, extensible subagents, and coding agents capable of broader work. Brian specifically wants Codex to use computers in novel ways and become more trusted, so small development shops and other teams can access capabilities associated with top-tier firms.