What is Codex?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
Codex is OpenAI’s agent product for coding and related project work. It combines model capabilities with a working environment in which the agent can inspect files, make changes and run tools. An agent is software that chooses actions toward a goal and uses feedback to decide what to do next.
Codex is not simply another name for every GPT model. A model is the trained component producing responses and action proposals. Codex names the product and agent experience around those capabilities. A model identifier containing “Codex” is a separate use of the name, so the context in which it appears matters.
Consider asking for a small expense-tracking application. The agent may help clarify requirements, create files, run the program and respond to failures. Requirements are statements of what the product must do. A useful assignment would define what expenses can be entered, how they are saved and which actions remain outside the project’s scope.
The environment and permissions determine what the agent can actually do. A terminal is a text-based place to run commands. A graphical application offers visible controls and panels. Remote execution means work takes place on another computer rather than the user’s own machine. These are working arrangements, not independent kinds of intelligence.
A repository is a project’s collection of files and recorded changes. Codex can work within such project structures, but its access should match the task. Reading a repository, editing it, merging changes and deploying a live version are distinct activities. They should not be treated as one implied authorization.
Whether planning or building, connect your instructions to observable behavior. Tests check whether expected behavior occurs; review examines the changes and their consequences. A claim of successful completion should be supported by what actually ran and what was inspected.
A healthcare team using Codex to build a referral-tracking prototype could begin with an agreed requirement: every unresolved fictional case shows its blocker, owner and last update. The team would then inspect the implementation and test missing information, duplicate records and interrupted saves. This is a practical extension of the course’s build-and-test approach, not evidence of clinical readiness. The person directing the agent still needs to decide what should be built, what is excluded and what must be demonstrated before real users depend on it.
Source: OpenAI — Codex Overview.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy