$ today_in_10_seconds
OpenAI shipped computer-use agents for healthcare that operate EHRs and payer portals the way a person does: reading the screen, clicking, typing, and switching apps. They don't wait for an API to exist. CVS and Cigna already run them in production, so the integration bottleneck healthcare has complained about for ten years now has a workaround.
$ lead_story
Advancing computer-use agents for healthcare teams
GPT-6 Astra and GPT-6 Luna now power agents that run multi-step work inside the software care teams already use. They're available through the API, ChatGPT for Healthcare, Codex, and ChatGPT for Clinicians. In physician-designed EHR inbox evals, Astra synthesized chart history to draft caregiver instructions and chart notes, and reviewers rated 99% of its actions clinically safe across 325 tasks. The design pattern matters more than the score: the agent drafts but doesn't send, pauses when chart evidence is ambiguous, and logs every click. Human oversight is built into the action layer from the start. For builders, 'there's no API for that' stops being a blocker in one of the most API-hostile industries there is.
// why_it_matters: GUI agents just turned healthcare's legacy software from a moat into an interface.
$ also_worth_attention
More ways to move care forward with ChatGPT for Healthcare
OpenAI added 20+ plugins, usable in ChatGPT and inside EHR workflows via ChartGPT. They cover clinical reference (UpToDate Expert AI, IMO Health's MCP server), interoperability (Redox, Metriport, Elation), and revenue-cycle systems (Availity, Stedi, TriZetto). OpenAI is running two strategies at once: structured tool calls where integrations exist, and computer use where they don't. Most serious agent stacks end up with that hybrid.
link →Clinical search got measurably better
In a blind evaluation of 1,300+ multi-turn clinical reviews, physicians rated the upgraded clinical search higher on all eight dimensions. Factual correctness rose from 93.4% to 97.6%, and clinical safety rose from 89.4% to 95.5%. For agents, retrieval quality comes before every action, so a better grounding layer makes everything downstream safer to automate.
link →The unit economics: Luna at 8.5¢ a task
On Stanford's HealthAdminBench, which covers prior authorizations, claim appeals, and equipment orders, Astra scored 94.3% of rubric points. The smaller Luna scored 91.8% at 8.5 cents per task, about 98% cheaper. Most high-volume back offices will happily give up 2.5 points of quality for a ~50x cost cut: run routine cases on the small model and send the hard ones up.
link →
$ bigger_picture
The trend behind the headlines
Health systems have spent years waiting for EHR vendors and clearinghouses to expose usable APIs. Computer use skips the wait. CVS cleared a multi-month document backlog over a single weekend. Cigna went from concept to production in under eight weeks, replacing care managers' manual navigation across 10+ systems. The losers are outsourcing shops and thin point-solution startups whose main value was 'we integrated with the ugly system.' The winners hold proprietary data, regulatory clearance, or the governance layer of audit trails, pause-on-ambiguity, and draft-not-send that gets an agent past a hospital risk committee.
CVS Health CUA Agent Kit
CVS's open-source package, built on OpenAI's Agents SDK, for deploying computer-use agents as a repeatable process. It's worth reading even if you never touch healthcare. It's a production governance blueprint from an enterprise running agents with human oversight and screenshot verification at the scale of tens of thousands of daily consults.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime