OpenAI updated its Codex coding tool this week: users can now configure AI with roles like "code analyst" and "reviewer" via TOML config files (a simple text format), and specify which tasks trigger each role and which model it calls. We think this matters—AI agents are moving from "one-size-fits-all assistant" to "configurable virtual team," and Codex's three-layer config (default settings, role definitions, dispatch rules) is the canonical blueprint for this approach.
What this is
Codex is OpenAI's AI coding assistant. The latest version introduces a three-layer configuration structure:
· Entry settings (config.toml): defines default model and reasoning intensity;
· Role definitions (TOML files in the agents directory): each file is a "position"—e.g., code_explorer (read-only tracing of code call chains), reviewer (fixed-model code risk checks);
· Dispatch rules (AGENTS.md): specifies when each role is called and what model it uses.
Simply put: previously you asked AI a question, it answered with one model; now you can have AI act like a company—dispatch an "analyst" to survey the scene first, then a "reviewer" to gatekeep—and each role can run a different model.
Industry view
Supporters see this as a sign of agent maturity—configuring AI like employees (positions, responsibilities, tool permissions) is the only way to truly land AI in enterprise workflows.
Critics aren't quiet. First, Codex's config still requires hand-writing TOML files, putting it out of reach for non-developers—proof that AI coding tools haven't reached "out-of-the-box" yet. Second, Anthropic's Claude Code has shipped similar custom agent capabilities for a while; OpenAI looks more like catching up than leading. Third, the "every vendor has their own config file" situation shows the agent ecosystem still lacks a unified standard—whoever hides the complexity wins non-technical users.
Impact on regular people
· For enterprise IT: start asking whether "AI employees" need job descriptions like human employees—but short-term, this stays an internal R&D tool.
· For working professionals: technical roles should pay attention; new tools mean shifting skill demands. Non-technical roles can ignore it for now.
· For consumer market: consumer AI products are still far from "configurable division of labor"—what regular users get is still a chat box.