The fluorescent-lit data centers of Fortune 500 companies hide a secret that IT executives would rather not discuss: Somewhere in the basement, or perhaps a regional office in Cleveland, legacy mainframes are still running payroll. Insurance claims still route through terminal applications that predate the graphical user interface. And somewhere, someone is manually typing data from one green-screen system into another because the two simply cannot talk to each other.
Coasty, a small San Francisco outfit backed by Y Combinator, thinks it has found a way in.
The startup's pitch is deceptively simple. Rather than trying to modernize these ancient systems—a task that has broken budgets and careers for decades—Coasty built an AI agent that interacts with them the way a human would. It watches the screen. It moves the mouse. It types. And according to the company, it can do this reliably enough to automate workflows that have resisted every previous attempt.
Whether that claim holds up in production is another matter entirely. But the underlying problem Coasty is chasing? That one's real enough.
The Ghost in the Machine
Large organizations run on software archaeology. Beneath the sleek customer-facing apps and cloud dashboards lie thousands of workflows that depend on systems no one quite knows how to replace. Maybe it's a mainframe handling state benefits, maybe it's a Citrix session running some custom accounting tool from 1987, maybe it's just Excel macros daisy-chained together by someone who retired eight years ago.
Traditional robotic process automation—RPA, in the jargon—tries to navigate this mess by identifying specific UI elements: click this button, type in that field, wait for this window. It works until someone updates the interface, at which point the whole automation collapses. And for systems that lack modern APIs entirely? The usual answer is a resigned shrug and a temp worker.
Coasty's founders, Prateek Jannu and Nitish Kovuru, are betting there's a better approach. Their agent doesn't look for buttons or fields. It takes screenshots, analyzes what's on the screen, and executes actions as raw pixel coordinates. No UI selectors to break. No APIs required. Just vision and mouse control, the same toolkit available to any human operator.
The technology runs inside isolated virtual machines—Linux or Windows, spun up on demand—that Coasty provisions in the cloud. According to the company's Y Combinator launch materials from mid-2026, the system handles "mainframes, Citrix, virtual desktops, and terminal apps" without requiring code changes to the underlying software. Which is precisely the point, since modifying that software is often impossible.
Benchmarks and Reality
Coasty's marketing leans heavily on a single number: 82.81%. That's the company's reported score on OSWorld, an academic benchmark designed to measure how well AI agents complete multi-step desktop tasks. For comparison, OpenAI's Operator—a computer-use agent the company detailed in early 2025—scored 38.1% on the same test.
More than double. Impressive, if the comparison is fair.
It might not be. OpenAI's published result from January 2025 is now over eighteen months old, and the computer-use agent field has been moving fast. Multiple vendors and open-source projects have iterated on the core vision-and-action model since then. OSWorld itself released a second, harder version of the benchmark in mid-2026, with longer and more complex tasks; the best-performing system on that version—Claude Opus 4.8 with extended reasoning—completed just 20.6% of workflows at the 500-step mark.
Coasty's verified score appears on the original OSWorld v1 leaderboard, though the company hasn't disclosed which underlying language model powers its agent. And anyone who has worked in enterprise IT knows that benchmark performance and production reliability are not the same animal. The gap between "works in the lab" and "works when the CFO is screaming about month-end close" is where many automation projects go to die.
Still, 82.81% is 82.81%. The company reports it has "paying customers on four continents" and "a few thousand active users," though it has not named any enterprise clients publicly. LinkedIn pegs the employee count at 2–10 people as of August 2026. Small team, big ambitions.
Under the Hood
For developers, Coasty exposes a straightforward HTTP API. Send a screenshot to the /v1/predict endpoint, get back an action: click at coordinates (412, 237), type "invoice-2847", scroll down 150 pixels. The /v1/sessions endpoint manages multi-step workflows, capturing full state and screenshot history as it goes. For complete end-to-end automations, the /v1/runs endpoint orchestrates everything from start to finish at five cents per step.
The system includes mid-run error recovery and output verification—addressing what Jannu and Kovuru describe as the Achilles' heel of traditional RPA. A selector-based bot will cheerfully click the wrong button, type into the wrong field, and submit garbage data without ever realizing something went wrong. Coasty's approach, the founders argue, builds in checks: Did that action produce the expected result? If not, try something else.
Each agent run generates what the company calls a "replayable audit trail," complete with hashed screenshots. For compliance teams in regulated industries, that audit capability is potentially more valuable than the automation itself. When a regulator asks what happened to transaction 4827-B, having a pixel-perfect record of every screen and every action is the kind of thing that makes legal departments happy.
Coasty offers three deployment modes, each targeting different use cases and security postures. Cloud-hosted VMs run in isolation—actual virtual machines, the company stresses, not containers in a shared pool. There's also a desktop application, open-sourced and built on Electron, that lets agents control local Windows or macOS machines for organizations that won't let workloads leave the building. And a feature the company calls "Agent Swarms," launched earlier this year, spins up multiple parallel VMs to divide batch work: processing a thousand invoices, say, or fulfilling a flood of e-commerce orders.
The startup has also published open-source repositories positioned as starting points for developers building their own agents. One, open-computer-use, provides core infrastructure; another, open-cowork, supports bring-your-own-key setups where customers pay their language model provider directly and bypass Coasty's usage fees.
What It Costs
The pricing model is pay-per-action, which means costs scale directly with usage. Managed inference—where Coasty provides both the infrastructure and the model—runs five cents per prediction, four cents per session step, and five cents per agent run step on the current model versions. Virtual machines bill at five cents per hour for Linux, nine cents for Windows, with a penny-per-hour standby rate when stopped.
For organizations that already have enterprise contracts with Anthropic or OpenAI, the bring-your-own-key option waives Coasty's usage fees, though customers still pay their model provider. The company offers free sandbox keys for testing, and the desktop application ships as a direct download for Windows and macOS. API access requires a request through the Coasty website.
Whether that pricing makes sense depends entirely on the workflow. Five cents per step adds up quickly if you're automating something simple. But if the alternative is paying a human $25 an hour to manually shuttle data between systems—and many enterprises are doing exactly that—the math tips in Coasty's favor pretty fast.
The Bigger Picture
Coasty's launch comes at a moment when legacy IT vendors are scrambling to bolt AI onto their existing platforms. Tools like Automic and Control-M, which have been orchestrating batch jobs since before "cloud" meant anything other than weather, are integrating AI agent interfaces and model context protocols. UiPath, which built an RPA empire on selector-based automation, shipped an orchestration layer mid-last year specifically designed to manage hybrid human-agent workflows.
The underlying question is whether pixel-level automation can actually deliver the kind of reliability that enterprises demand. Benchmarks measure task completion under controlled conditions, often with clean test environments and well-defined success criteria. Production systems need uptime, graceful failure handling, and audit trails that satisfy not just IT but also legal, compliance, and sometimes federal regulators.
Coasty's emphasis on verification and mid-run recovery suggests the founders understand those requirements. Jannu, a Stanford dropout who previously studied at Purdue, and Kovuru, a Columbia CS graduate, both have backgrounds in enterprise systems. They're betting that IT leaders will pay for reliability over raw benchmark scores—and that vision-based agents will prove more robust than the selector-based tools they're trying to replace.
Maybe they're right. Or maybe they'll discover that automating legacy systems is hard for reasons that have nothing to do with technology: organizational inertia, risk-averse IT departments, the eternal question of who gets blamed when the automation breaks something expensive.
For now, the product is available, and the company has published tutorials covering mainframe automation, insurance claims processing, and real estate workflows—use cases the company identifies as areas where manual workarounds persist. The pitch to IT leaders managing constrained budgets and aging infrastructure is direct: Automate the systems that can't be replaced.
Whether enterprises bite is the next chapter. The technology exists. The problem is real. What remains to be seen is whether anyone wants to be the first to trust a startup with two to ten employees to automate their most critical, most fragile systems.
That's a bet some CIO is going to have to make.
