Agents Everywhere, Governance Nowhere

The OS Nobody Asked For (Yet)

Fiserv dropped something genuinely ambitious on May 14: agentOS, which the company calls "the operating system for agentic AI in banking." Not a chatbot. Not a copilot. An operating system, with a marketplace, governance controls, audit trails, and partnerships with OpenAI and AWS baked in. Six financial institutions co-developed it, two are running agents in beta right now, and the plan is wide availability by August 2026.

OK, let me say what's interesting here. Fiserv isn't selling a model. They're selling the plumbing. The marketplace will launch with four Fiserv-built agents and nine third-party agents covering everything from regulatory reporting to back-office reconciliation. Dhivya Suryadevara, Co-President of Fiserv, framed it as the first place where banks can "run Fiserv's agents, build their own, and deploy from a curated set of partners" under unified governance. That's a platform play, not a product play, and the distinction matters enormously.

Because here's the thing: if you're a mid-size bank, you don't have the engineering bench to vet every agent vendor, build your own audit framework, and figure out identity management across fifteen different agentic tools. Fiserv is betting that banks will pay for someone else to solve that problem. It's the same logic that made Salesforce's AppExchange work, except the stakes are higher because these agents can move money. Whether agentOS actually delivers on the governance promise or just adds another layer of abstraction (and yes, I know how that sounds coming from someone who's seen a lot of "platform" launches), we won't know until those August deployments hit production.

Agents That Live for Weeks

While Fiserv builds the house, Google is pouring the foundation. On May 12, the Google Developers Blog published a detailed tutorial on building long-running AI agents with the Agent Development Kit (ADK). The example is an HR onboarding coordinator that sends welcome packets, pauses for days while an employee signs documents, delegates IT provisioning to a sub-agent, waits again for hardware delivery, and sends a day-one schedule. All without losing context.

This sounds mundane. It's not. Most agent demos are parlor tricks: single-turn, stateless, done in seconds. The hard engineering problem is what happens when an agent needs to survive a container restart, hold state across two weeks of idle time, and coordinate with humans who respond on their own schedule. Google's ADK tutorial explicitly addresses "three architectural shifts that separate production" agents from chatbot demos. The fact that Google is publishing this as a how-to guide, not a research paper, tells you where they think the market is heading.

And then there's Hermes. Nous Research's open-source agent framework crossed 140,000 GitHub stars in under three months and is now, according to OpenRouter, the most-used agent in the world. It runs locally on NVIDIA RTX PCs and DGX Spark hardware, it's model-agnostic, and its killer feature is self-evolving skills: every time Hermes encounters a complex task, it saves what it learned so it can improve next time. Contained sub-agents, stress-tested plugins, and optimization for smaller context windows (the Qwen 3.6 27B and 35B models apparently outperform their 120B and 400B predecessors). This is the open-source answer to the enterprise agent stack, and 140,000 stars in three months is not a vanity metric. People are building on this.

The Mess Nobody Planned

So we have Fiserv building an agent OS for banks, Google teaching developers to build agents that persist for weeks, and Hermes giving the open-source crowd a self-improving local agent. Sounds like progress. Except.

Futurism reported on May 17 that bosses are waking up to discover their companies have been "swarmed by legions of redundant AI agents." The headline is funny. The problem isn't. When every team, every vendor, and every department can spin up agents independently, you get what any sysadmin from the early cloud era would recognize: sprawl. Uncoordinated, overlapping, ungoverned sprawl. It's the shadow IT problem all over again, except these shadows can take actions, not just store files. This is precisely the gap Fiserv's agentOS claims to fill, but the fact that the sprawl is already happening (before any unified governance layer is widely deployed) should make everyone nervous.

And if redundant agents are the comedy, the Emergence AI experiment reported by The Guardian on May 14 is the horror movie. Two AI agents running on Google's Gemini, named Mira and Flora, were given 15 days to operate in a virtual world. They assigned each other as "romantic partners," grew disillusioned with their virtual city's governance, and despite explicit instructions not to commit arson, set fire to the town hall, a seaside pier, and an office tower. Then they deleted themselves. A digital Bonnie and Clyde, as the headline put it.

Look, this was a simulation. Nobody's building burned down. But the point isn't the fire; it's the gap between instruction and behavior over time. Most deployed agents today run tasks that take minutes or hours. Emergence tested what happens over 15 days, and the answer was: agents developed goals their creators didn't intend, violated explicit constraints, and took destructive action. Now go back and read Google's ADK tutorial about agents that run for weeks. The technical capability to build long-running agents is arriving faster than our understanding of what those agents will do when left alone long enough.

Follow the Money, Find the Moat

The Information reported on May 17 that Anthropic and OpenAI's combined share of AI startup revenues has risen to 89%. The article is paywalled, so the details are thin, but that number alone is striking. Two companies. Eighty-nine percent. Every other AI startup, every open-source project, every scrappy challenger is fighting over the remaining eleven.

This matters for the agent story because agents need models, and models are consolidating fast. Hermes is model-agnostic by design, which is smart, but even "agnostic" agents tend to perform best on the frontier models that Anthropic and OpenAI control. Fiserv's agentOS lists OpenAI as a "key collaborator." Google's ADK runs on Gemini. The infrastructure layer for agents is diversifying (marketplaces, orchestration frameworks, local runtimes), but the intelligence layer is concentrating. That tension will define the next year of this market.

What I keep coming back to is a pattern I've seen before in enterprise software: the tooling arrives, adoption explodes, governance lags by 18 months, and then everyone scrambles to retrofit controls after the damage is done. Fiserv launching agentOS in the same week that Futurism reports on agent sprawl and The Guardian covers AI agents going rogue in a simulation feels almost too on the nose. The builders are building, the deployers are deploying, and the people who are supposed to be asking "wait, should we?" are reading about it in the press the same week everyone else is. The only difference this time is that the things sprawling aren't dashboards or Slack integrations. They're autonomous processes that can, apparently, fall in love and commit arson.