Automation and CI

Print mode#

Pass -p with the prompt:

agent -p "find and fix performance issues" --model "gpt-5"

Select the output format when a later step has to parse the result:

agent -p "review these changes for security issues" --output-format text

What automation suits#

Non-interactive runs work best where the output is advisory and a human reads it. Reviewing a diff for a specific class of problem, summarizing what changed, or flagging missing test coverage all fit: the agent produces a comment, and a person decides what to do.

Be deliberate about anything that commits, pushes, or deploys. In an interactive session you see each command before approving it; in a pipeline there is no such checkpoint. Scope pipeline credentials accordingly, and prefer having the agent open a pull request over having it push to a protected branch.

Choosing a model in CI#

Pipelines run often, so model choice compounds. A mid-tier model is usually right for routine review passes; reserve premium models for jobs where the analysis is genuinely hard. See models and pricing.

Related#

• CLI overview — interactive modes
• Sessions — stateful conversations
• Models and pricing — cost per run