OpenAI presented new results and deployments of AI agents that operate software interfaces directly on November 16, 2026, with examples from healthcare organizations' administrative processes. The announcement describes collaborations with CVS Health, Cigna Healthcare, Abridge, and Ambience. In the same context, CVS Health published CUA Agent Kit, a project that converts successful agent runs into inspectable automations subsequently executed without model calls.

The connection between the two components is work across applications whose APIs do not cover every necessary operation. A computer-use agent can navigate pages, enter information, and move between systems through the interfaces available to teams. CVS Health's project separates discovering these steps from subsequently repeating them through code, for workflows with stable, reproducible stages.

Four support team members wearing headsets with microphones in an office
Mikhail Nilov

From AI-assisted navigation to execution through code

CVS Health presents CUA Agent Kit as a starting point for automating browser and Windows activities. In the initial stage, an OpenAI agent follows a task description written in ordinary language. The run records screenshots, actions, targets, and execution traces, creating evidence of the steps performed.

After successful runs have been obtained, a Codex procedure uses that evidence to build or update a deterministic workflow. For browsers, the project produces Playwright code that identifies interface elements through their meaning. For Windows applications, it uses accessible controls and pywinauto code. The team can inspect and maintain these operations as ordinary software.

The documentation specifies that deterministic runs make no model calls and do not require the API key used during discovery. Access to the target applications remains necessary to perform the operations. When a workflow changes or breaks, Codex can use the recordings for diagnosis and code updates within a process that includes engineer review, testing, and outcome verification.

The project is publicly available under the Apache 2.0 license. Documented requirements include Python 3.12 or later, the uv tool, and access to the website or environment being automated. The package includes a local browser demonstration application. Its Windows example saves a Unicode document in Notepad and verifies both the saved file and its contents reopened in the editor.

CUA Agent Kit stageTechnology's roleRetained evidence
Workflow discoveryAI agent follows instructions and operates the interfaceScreenshots, actions, targets, and execution traces
CodificationCodex builds or updates deterministic stepsInspectable code and a verified outcome
Repeated executionPlaywright or pywinauto, without model callsRun records and final-state checks

The implementations OpenAI describes

For CVS Health, OpenAI mentions sending documents, processing referrals, and supporting specialist consultations across clinical systems. The company describes a production workflow that cleared several months of accumulated documents over a weekend. Another workflow performs software steps for tens of thousands of consultations daily, according to OpenAI's announcement.

These results are presented as examples from that organization. CVS Health's implementation uses validation and human oversight, including screenshots and action records to verify application state. Data remains within enterprise-controlled infrastructure, according to OpenAI's account, governed by CVS Health's internal AI governance policies.

For Cigna Healthcare, the announcement describes bringing together information about benefits, providers, appointments, and previous interactions to support teams contacting members. OpenAI states that case managers previously navigated more than ten clinical and administrative systems. The agent's interface operation is presented as a way to gather information from those systems.

Abridge is described as developing agents to coordinate care plans across applications, including systems without APIs. Its inpatient example is a planned use case. Ambience presents a demonstration in which an agent prepares an insurer request, asks for clarification when the record is ambiguous, and waits for approval before submitting. These stages differ from the CVS Health workflows described as running in production.

Benchmark results are reported by the vendor

OpenAI also published results for HealthAdminBench, a benchmark of healthcare administrative tasks developed by Stanford researchers and their collaborators. According to OpenAI, GPT-6 Astra earned 94.3% of the available rubric points, while GPT-6 Luna earned 91.8%. Both models were evaluated at their maximum reasoning effort.

The company reports a cost of 8.5 US cents per task for Luna, 98% lower than Astra in that evaluation. The score percentages describe benchmark criteria rather than the share of a business's real operations completed without oversight. They are results reported by OpenAI on an external benchmark, rather than measurements performed by WebHat.

The HealthAdminBench project describes tasks involving prior authorizations, medical equipment orders, and appeals in environments with graphical interfaces. Evaluation combines checks of application state with information accuracy criteria. This structure examines operations carried out across several steps, extending beyond the assessment of an isolated textual response.

Access and data controls are defined separately

OpenAI identifies availability through its API, ChatGPT for Healthcare, and ChatGPT for Clinicians. For existing customers, workspace administrators manage the activation of computer use in ChatGPT Work and Codex. The individual free access route described for ChatGPT for Clinicians is aimed at US clinicians; the announcement does not present it as general free access for every business.

ChatGPT for Healthcare includes role-based access controls and audit logs, according to OpenAI. Organizations can configure accessible applications and the steps requiring approval. For workflows involving protected health information under HIPAA, the company requires eligible services covered by its Business Associate Agreement and the necessary security and retention settings.

CUA Agent Kit's documentation separately explains that discovery sends instructions, screenshots, and action history to the model. Local recordings can contain visible data, action arguments, session identifiers, or authentication information. Deterministic execution without a model and AI-based discovery therefore have distinct data flows, even though they form part of the same automation.