codex exec runs the agent without a TUI - it is the entry point to non-interactive work, scripts and CI. The key design decision here is how the output is arranged. Progress goes to stderr, and the agent's final message to stdout. This separation is not cosmetic: it lets you safely pipe the output without mixing the result with the course of reasoning. A script gets a clean result on stdout, while all the progress noise stays on stderr, where it is expected.

Basic use reads like an ordinary command embedded in the shell. It helps to see examples once. Below are a one-off request, redirecting the result via tee into a file and an ephemeral run that leaves no session. Precisely because the final result goes to stdout, it can be passed on like the output of any other tool - written to a file, passed to the next command, processed. exec turns the agent into a link of a pipeline rather than a separate application.

For programmatic parsing there is a machine-readable mode. The --json flag emits a stream of events that can be processed via jq: instead of human-readable text you get structured events of the work's course. This is the basis for integration into scripts that need not just the final message but visibility of what is happening - which steps the agent took, what it called, how it ended. A machine-readable stream is what makes exec suitable for automation, not only for a one-off run.

Structured output is the most powerful part of exec for integration. It helps to see how it works once. Below is a JSON Schema describing the expected shape of the result: for example, the number of risks and an array of their descriptions. By setting a schema, you get from the agent not free text but data of a defined structure that can be parsed at once without parsing prose. This turns the agent's output from something to interpret into something that can be consumed programmatically.

The schema is attached with explicit flags, and the result is put into a file. Below is a request with --output-schema pointing to a schema file and --output-last-message saving the final result to a file. The bundle is simple and reliable: the schema sets the shape, the last message is saved separately from the progress stream. After the run a script reads a ready JSON of the defined structure rather than reading the result out of mixed output, guessing where progress ends and the answer begins.