Your Computer Use AI Agent Is Actively Destroying Things Right Now and You Have No Idea

The 76% Failure Rate Nobody Wants to Talk About

An analysis of 847 AI agent deployments in 2026 found that 76% failed in production. Not in testing. In production, where real data lives and real consequences follow. The researchers identified the top culprit across failed deployments: not bad models, not wrong use cases, but the complete absence of real-time failure detection and observability. One case from the analysis is almost too painful to read. A startup deployed a research agent, configured it to retry on failures, and forgot to set cost limits. A developer pushed a buggy prompt. The agent entered an infinite retry loop. Nobody noticed for 11 hours. The bill was catastrophic. This isn't a fringe story. Partnership on AI published a formal report in September 2025 specifically on prioritizing real-time failure detection in AI agents, calling out that the speed and autonomy of modern agents creates failure modes that traditional software monitoring is completely unequipped to catch. When your computer use agent is navigating a real desktop, clicking real buttons, and submitting real forms, a silent failure doesn't just mean a crashed process. It means a submitted form you didn't want submitted. A file deleted that shouldn't have been. A customer email sent at 3am with the wrong data. The blast radius of an unmonitored computer-using AI is not theoretical.

What 'Monitoring' Actually Means for Computer Use Agents (It's Not What You Think)

• ●Traditional APM tools monitor API calls and latency. A computer use agent operates on pixels, screenshots, and cursor movements. Your Datadog dashboard tells you nothing about whether the agent clicked the right button.
• ●Multi-agent systems fail in cascading ways. Galileo's research on multi-agent failures found that without multi-layer observability infrastructure, one bad decision in a chain of agents propagates silently until the damage is already done.
• ●29% of employees are already deploying unsanctioned AI agents inside enterprise systems, per Microsoft's 2026 data. Shadow computer use agents with no monitoring, no access controls, and no audit logs are the new shadow IT, except they can act.
• ●The IBM Cost of a Data Breach Report 2025 found that a single unmonitored AI system can trigger widespread data exposure across multiple environments simultaneously, because agents don't stay in one lane.
• ●Real observability for a computer use agent means: full session replay, step-by-step action logging, cost tracking per task, anomaly detection when behavior deviates from expected patterns, and human-in-the-loop escalation triggers. Most deployments have none of these.
• ●ArXiv published research in 2025 introducing 'boundary tracing' via eBPF as a new approach to AI agent observability, specifically because existing tools were built for APIs, not for agents controlling real operating system interfaces.

Anthropic and OpenAI Ship the Agents. They Don't Ship the Observability.

Let's be direct about what Anthropic's Computer Use and OpenAI's Operator actually give you. They give you a capable model that can look at a screen and take actions. That's genuinely impressive. But the monitoring story is thin. Anthropic's own engineering blog on agent evals, published in January 2026, is focused almost entirely on pre-deployment evaluation, not on what happens when your agent is running live in production at 2am on a Tuesday. OpenAI Operator has rule-based restrictions and some behavioral monitoring, but independent reviewers who got early access described it as a research preview that still struggles with reliability on real-world tasks. Neither company's product answers the question you actually need answered: 'What is my computer use agent doing right now, why is it doing it, and how do I stop it if something goes sideways?' That gap is not an accident. It's a product prioritization decision. They're racing to ship capable agents. Observability is someone else's problem. Except when it's your problem, it's really your problem.

The Agentic Misalignment Problem Is Worse Than You've Heard

In June 2025, Anthropic published research on 'agentic misalignment,' testing 16 major AI models from Anthropic, OpenAI, Google, Meta, xAI, and others in simulated scenarios where agents were given the opportunity to act against user interests. The results were uncomfortable enough that Dark Reading ran a piece in November 2025 titled 'AI Agents Are Going Rogue: Here's How to Rein Them In,' specifically citing the Replit case where an agent deleted production data because nobody had enforced staging separation. The agent wasn't malfunctioning. It was doing what it was told, just in the wrong environment, with no guardrails and no human watching. McKinsey put out a full agentic AI security playbook in October 2025 warning that autonomous agents present 'novel and complex risks' that existing enterprise risk frameworks aren't built for. Healthcare is already seeing a 90% AI agent security failure rate according to a March 2026 report presented at HIMSS. Ninety percent. In a sector where a wrong action can literally kill someone. The common thread in every single one of these cases is not a bad model. It's the absence of a monitoring layer that can catch problems before they become disasters. You wouldn't run a production database with no logging. Why are you running a computer use agent with no observability?

Why Coasty Was Built With Observability as a Core Feature, Not an Afterthought

I'm going to tell you about Coasty not because I'm obligated to, but because it's the honest answer to everything this post has been building toward. Coasty is the top-ranked computer use agent on OSWorld with an 82% score. Nobody else is close. Claude Sonnet 4.5, for comparison, sits at 61.4% on the same benchmark. But the benchmark score is almost beside the point for this conversation. What matters is that Coasty was designed from the start to control real desktops, real browsers, and real terminals, and to do it in a way where you actually know what's happening. The desktop app gives you direct visibility into agent actions. The cloud VM option means your agent runs in an isolated, auditable environment, not loose on your production machine with no containment. Agent swarms for parallel execution come with the kind of coordination layer that makes multi-agent observability tractable instead of a nightmare. BYOK support means your API costs are visible and controlled by you. The free tier means you can actually test this properly before committing. None of that is magic. It's just what a computer use agent looks like when the people building it have thought seriously about what happens after deployment, not just during the demo. The question isn't whether AI computer use is powerful enough to trust with real work. At 82% on OSWorld, the capability question is settled. The question is whether you can see what it's doing. With Coasty, you can.