Coasty turns browser automation into a revenue-cycle operating system

Coasty, a computer-use agent platform ranked #1 on OSWorld at 85.6% task completion, is not selling itself as a general-purpose automation tool. It is selling itself as a revenue-cycle operating system for healthcare billing teams. The pitch is precise: run the 40-login morning, the swivel-chair data entry, the hold-music workflow, the denials you learn about 30 days later. Run them on a loop that observes, acts, verifies, and recovers, just like a careful biller, but at machine speed.

The distinction matters. Plenty of AI agent companies demo a browser doing impressive things. Coasty is naming the specific pain points — Availity, UHC/Optum portals, Epic, athenahealth, Waystar, Trizetto — and the specific workflows: eligibility verification, prior auth submission, claim scrubbing, denial triage, ERA posting. The company’s website reads less like a product page and more like a day-in-the-life of a medical biller, complete with the password spreadsheet and the 15-minute session timeout.

The core insight is that healthcare billing is a screen-based industry that has never gotten its API act together. Payer portals ship without APIs. Clearinghouses charge per-connection fees. Legacy practice management systems predate modern integration patterns. The result is an industry that runs on human eyeballs and muscle memory: tab over to this field, copy that number, paste it here, wait for the page to load, check the response, move to the next portal. Coasty’s bet is that a vision-based agent can replace that human-in-the-loop for the parts of the job that are not judgment work, while leaving the judgment calls — coding, clinical decisions, payer-facing submissions — with the people.

The company claims 85.6% on OSWorld, a benchmark that tests agents on real computer tasks across 369 scenarios. The second-place entry, Pointer Agent with Opus 4.7, scored 83.6%. Coasty’s public model scored 82.81% on the independently verified OSWorld leaderboard. The company publishes traces on GitHub. The benchmark performance is credible, but the harder test is whether the agent generalizes across the specific portals and edge cases that healthcare billing throws at it.

Coasty’s architecture is straightforward. The agent runs in a sandboxed virtual machine — Azure Container Instances or AWS EC2 — with a full Ubuntu desktop, Chrome, and a WebSocket bridge. It supports multiple models through Amazon Bedrock: Claude Sonnet 4, Claude 3.5 Sonnet, Claude 3 Haiku, Llama 3.2, and Amazon Nova. Users bring their own API keys. The agent sees the screen, reasons about what to do next, and executes clicks, keystrokes, scrolling, and drag-and-drop. Every action is logged with a screenshot. Gated steps — claim submissions, appeal filings, auth requests — pause for human approval before anything reaches a payer.

The security architecture is designed for healthcare compliance. PHI stays under the customer’s control. Coasty can deploy into a customer’s own VPC, behind their firewall, with their identity controls. Screenshots and logs are stored inside the customer’s environment. The only data that leaves is the inference call to the model provider, under the customer’s own agreement and BAA. The company offers BYOK encryption with AES-GCM for API keys. There is no Coasty data lake.

The business model is credit-based billing, with a free tier and plans from $9 to $100 per month. The desktop app — a cross-platform Electron overlay — lets agents run on the user’s own machine with four approval modes from full autonomy to manual review of every action. The platform also supports agent swarms, scheduled automation, and web research at three depth levels.

What is genuinely new here is not the agent technology. Computer-use agents have been a category since Anthropic’s Claude 3.5 Sonnet introduced computer use in October 2024. What is new is the vertical specificity. Coasty is not asking healthcare billing teams to figure out how to apply a general-purpose agent to their workflows. It is naming the workflows, the portals, the error codes, the denial buckets, the industry metrics. It is speaking the language of days in A/R, first-pass yield, initial denial rate, and cost to collect.

The company’s website cites industry figures: 3-4% of net revenue for cost to collect, initial denial rates of 10-15%, most denied claims never reworked. These are not Coasty customer measurements — the company is transparent about that — but they are the levers a revenue-cycle leader already tracks. The argument is that an agent that absorbs eligibility runs, status pulls, and edit queues collapses the marginal cost of each touch. Appeals that were quietly written off because an hour of biller time against a small balance lost money on paper become claims that get worked when drafting a CARC-coded appeal packet takes minutes of review.

The open question is whether the agent actually generalizes across the long tail of payer portals. The portals Coasty lists — Availity, UHC, Optum, Waystar, Trizetto, Change Healthcare, Medicaid portals — are the major ones, but healthcare billing involves dozens of payer-specific interfaces, each with its own quirks, session timeouts, and error messages. The OSWorld benchmark covers 369 tasks, but those tasks are not healthcare-specific. The real test is whether Coasty’s agent handles the portal that rearranges its menu on a Tuesday afternoon, or the fax-back auth that requires a human to physically print and scan a form.

Coasty’s answer is the observe-act-verify-recover loop. The agent reads the screen, acts through mouse and keyboard, verifies the result, and recovers when the portal does something unexpected. It is the same loop a careful biller runs. The question is whether the recovery mechanism is robust enough to handle the full range of portal pathologies without human intervention.

For AI builders, the lesson is that the most valuable applications of computer-use agents may not be the flashy demos — the agent that books a flight, the agent that fills out a job application — but the boring, high-volume, screen-based workflows that have resisted automation for decades because they lacked APIs. Healthcare billing is one. Insurance claims processing, mortgage underwriting, government benefits administration, logistics coordination: the list of industries that run on portals is long, and the portals are not getting APIs.

Coasty’s approach is to treat the screen as the interface, not as a problem to be solved with integration projects. It is a bet that vision-based agents have crossed the reliability threshold for workflows where the cost of a mistake is a re-do rather than a catastrophe. Gated steps handle the high-stakes submissions. The rest runs on the loop.

The company is not claiming to replace billing teams. It is claiming to absorb the parts of the job that are not judgment work, so the team can spend its time on the claims that need a human. That is a narrower pitch than “AI agent for everything,” but it is a pitch grounded in the actual economics of a revenue cycle where most denied claims are never reworked because the labor arithmetic does not pencil out. If the agent changes that arithmetic, the number that moves is not a benchmark score. It is days in A/R.