November 17, 2026AgentsAgent-OperableTool

OpenAI Puts Computer-Use Agents in the Clinic, Where the APIs Stop

Hospitals run on systems nobody built an API for. That is the opening OpenAI went after on Monday with computer-use agents for healthcare teams, built on GPT-6 Astra and aimed at the electronic health record inbox: referrals, documents, consults, prior authorizations. Announcement: https://openai.com/index/advancing-computer-use-for-healthcare-teams/

OpenAI's own numbers: 99% of 325 reviewed tasks were rated safe, and physicians rated the output very or extremely helpful in 93% of 646 assessments. Treat those as vendor evals. The partner list is more telling. CVS Health uses it where APIs do not reach, with data staying in CVS infrastructure. Cigna says its case managers used to navigate more than ten systems. Assort closes a referral step in Athena that the API cannot do, through Codex computer use. Abridge runs concurrent agents on discharge barriers.

The real product is the control pattern, and Ambience describes it best for prior authorization. The agent stops and asks when the chart is ambiguous. It waits for approval before it submits anything. Allowed actions are configurable, a human can pause it at any time, and every step goes into an audit log. Stop, ask, approve, log. That is not a nice-to-have in a clinic. It is the license to operate, and it is the same pattern showing up everywhere this month, from human gates in coding workflows to biometric approval in agent wallets.

It is available now in ChatGPT for Healthcare, ChatGPT for Clinicians and the API, with a BAA for HIPAA. A companion post adds 20-plus healthcare plugins: https://openai.com/index/more-ways-to-move-care-forward-with-chatgpt-for-healthcare/ . The bigger point is that computer use has found its job. Not replacing APIs, but filling the gaps between the systems that will never get one. Healthcare has more of those gaps than any industry, which makes it the hardest test and the best market at once.
← Previous
Attackers Now Run Patch-to-Exploit With No Human. Defenders Cut the Clock to 30 Days
Next β†’
Your Smart Orchestrator Is Only as Safe as Its Cheapest Subagent
← Back to all articles

Comments

Loading...
>_