Computer-use agent or API call: use Sume jobs, not clicks

When an agent needs to generate media, have it call Sume's API or hosted MCP tools instead of driving a web dashboard with computer use. API calls return structured results, carry an idempotency key, and can be capped before they spend; clicks cannot.

What changed at OpenAI

The OpenAI changelog for September 29, 2026 says computer use was added to the Agents API: "Agents can complete tasks in an OpenAI-hosted browser, with website access approvals and sign-in handled by your application." In other words, your code is responsible for approvals and sign-in when an agent browses.

Why an API call is the better tool here

A browser agent must find buttons, wait for pages and keep a login alive. An API call states the intent in one request. Sume's keys resolve workspace, owner and key metadata from the key itself, so the request needs no user id and no session. Do not send workspace_id, owner_user_id or user_id in request bodies.

Send one credential only: x-api-key or Authorization: Bearer. Sending both returns 401 unauthorized.

Gates that clicks do not have

Hosted MCP requires an idempotency_key on write and paid tools, supports dry_run=true to preview admission and cost without submitting, and enforces max_spend_usd when you provide it. OAuth sessions with mcp:read only see read-only tools; mutating and paid tools need mcp:write. There is no mcp:paid scope.

Results come back as jobs. Use jobs_wait and jobs_result rather than watching a page.

• Retry safely: the same idempotency_key deduplicates.
• Cap spend with max_spend_usd.
• Preview with dry_run=true before a burst.

When computer use still makes sense

Use it for tasks that have no API, such as a third-party site you must operate. For generation, upscaling, captions and trimming that Sume exposes as endpoints or tools, call those.