Where does a person approve a tool call?
The Python SDK can pause before a tool runs and resume after a decision. Your app needs to show what would change, check that the reviewer has permission, and enforce the decision before saving anything.
On paper, draw a proposed title change for a fictional workshop listing. The approval shows the old and new text. Rejection keeps the existing title. If approval arrives after another edit, ask the reviewer to inspect the new draft.
What should your app keep when it pauses?
The SDK can serialize a pending run so the app can resume it. Keep the saved state in trusted server storage and give the reviewer only the proposed change they are allowed to see.
Confirm an approval belongs to that exact pending change and consume it only once. After an interruption, check whether the write already happened before retrying. Resuming a saved run does not make writes run exactly once.
What can a run record contain?
Tracing records steps such as model and tool calls. In the Python SDK, tracing and capture of model and tool inputs and outputs are enabled by default. Check where traces go and disable sensitive-data capture or tracing before using private information.
A pending run and a trace are separate. Decide what each should retain and who may access it. Keep private records and secrets out of the paper exercise.
Does installing the SDK cover model use?
The Python package has an MIT license. Model, connected service and hosting charges are separate. Check the provider path and features you use, because support can differ between providers.
Budget for more than one model request if a run contains several model steps. Keep the app code, workflow settings and the formats of your stored state under your control.