How do I review Montage tool approval before an OpenAI Agents SDK run?
A remote tool being available to an SDK agent does not mean its effects were approved. Establish a pre-run review that validates the direct read path and documents the chosen execution boundary without initiating paid hosted execution or delegating approval decisions to the model. A pre-run review should leave hosted execution and retained-workspace effects pending. Direct local status can establish a useful baseline without implying that a paid agent run occurred or that every future tool action was approved.
Help me review Montage tool approval before an OpenAI Agents SDK run. State that the application is inspecting tool availability and approved status access before any agent run. Report the baseline and the exact later execution decision still required. No writes.Define the review-only phase
State that the application is inspecting tool availability and approved status access before any agent run. Leave all retained-workspace changes and hosted Responses execution outside this preliminary task. The review contract should explicitly leave paid agent execution outside this stage, preventing a tool-approval investigation from itself starting the workflow under review.
Inspect the selected interface
Review whether the application plans local MCPServerStreamableHttp or HostedMCPTool. Approval settings and authorization must be reviewed for that actual path rather than assumed equivalent between both arrangements. An interface inventory should identify the path whose approval semantics matter instead of mixing local SDK controls with hosted Responses configuration guidance.
Constrain proposed operations
Separate a status read from any suggested watchlist or audience mutation. A model’s ability to select a discovered tool is not the authority that approves its remote effect. The proposed operation’s intended effect is the basis for a future approval decision; discoverability alone supplies no authority to create retained workspace records.
Establish a direct baseline
Use the approved local connection to execute signals_status without inference. Inspect the real workspace while leaving any future hosted call explicitly unverified and unexecuted. The local baseline must remain labeled as local even if the application ultimately plans hosted execution, because auth and data boundaries differ.
Hand off the pending decision
Report the baseline and the exact later execution decision still required. Do not portray a pre-run review as a completed agent workflow or imply standing permission for all remote tools. A useful handoff names the still-pending execution and approval decisions rather than reporting a successful end-to-end agent run that never actually occurred.
Keep the boundaries clear.
- The proposed operation’s intended effect is the basis for a future approval decision; discoverability alone supplies no authority to create retained workspace records.
- A useful handoff names the still-pending execution and approval decisions rather than reporting a successful end-to-end agent run that never actually occurred.
Start with a read-only check: ask your agent to report actual source coverage before collecting records or exporting an audience.
Primary references
Documentation reviewed 2026-11-15. Product capabilities, client availability and plans can change. Third-party marks identify their products, not an affiliation.