Latest in AI
CVS CUA Agent Kit: Turn Computer-Use Discovery into Deterministic Workflows
The open-source kit separates learning a browser workflow from repeating it, with important limits around evidence and credentials.
CVS Health's CUA Agent Kit turns AI-discovered workflows into Playwright browser code or pywinauto Windows code. Codified runs need no model calls or model API key; ordinary target-application authentication still applies.
The useful question is where the workflow needs fresh judgment, and where its outcome can be specified and checked in advance.
What the kit changes
OpenAI highlighted the project in its November 16 healthcare computer-use announcement. The release's healthcare examples establish the context, but they do not make every workflow built from the repository clinically validated or ready for patient data.
The CVS repository includes a local browser sample and uses Codex to codify discovery evidence. This is a review of its published design, not a hands-on reliability test.
Consider a routine internal task that finds a record, checks its status and downloads a document. If the fields and decision rules are stable, repeated model reasoning may be unnecessary. If the operator must interpret an ambiguous request or decide whether two records describe the same person, that uncertainty needs its own handling rather than being hidden inside a recorded sequence.
Successful discovery is evidence to inspect
The project's Playwright codification skill tells Codex to verify the intended terminal state rather than trust a success message. It prefers multiple successful recordings and stable locators, such as accessible roles and labels, over replaying coordinates.
It also requires attention to popups, frames, loading states and changed prompts. When evidence does not establish a stable target or outcome, another discovery run is needed. The intended output is maintained automation code with checks, not a recording that somehow becomes reliable by itself.
For the document example, “the download button was clicked” is a weak acceptance criterion. A better one is that the expected document for the selected test record exists and the record was not modified. Write that outcome before running discovery, so a plausible-looking sequence cannot redefine success afterward.
Start with a disposable workflow
I would evaluate this approach with synthetic records and a narrow, reversible task before connecting it to real operations. Make the expected output easy to inspect, then compare the discovered behavior with the written requirement.
A useful test plan includes these cases:
- The expected record and document are present.
- The record exists but has no matching document.
- Two records share a similar name.
- The session expires before the task finishes.
- The workflow is started again after a successful run.
These are proposed checks, not tests performed for this article. Their purpose is to expose where the process needs a clear stop, a retry or a human decision. A fast happy-path run gives little information about those boundaries.
Treat credentials and traces as sensitive
The README warns that discovery sends prompts, supplied credentials, screenshots and action history to the model endpoint. Traces may retain cookies, tokens and network content, and redaction is incomplete. Optional uploads do not sanitize artifacts; keep them out of commits and public issues.
Before using real accounts, decide which endpoint may receive the data, who can inspect the evidence and when it will be deleted. A model-free final run does not undo earlier disclosure during discovery. Keep that review separate from whether the automation completes its task.
Where deterministic automation still needs judgment
The code will need maintenance when the application or intended outcome changes. The codification skill calls for renewed evidence, checks and authorized execution against the test environment after changes; it does not promise permanent compatibility.
I think the strongest use case is a repetitive process whose completion can be verified cheaply. The weaker fit is a process where almost every run requires a new interpretation. In either case, keep the cost comparison honest: include discovery, code review, exception handling and maintenance alongside inference costs.
The valuable output is a workflow you can inspect and verify. Removing repeated model calls helps only when the resulting code still performs the right job.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help developer-tool teams turn technical workflows and product changes into clear, practical content.