All views expressed here are my own and do not represent the views of my employer.
I had a conversation last month with the CTO of a top-20 US bank. She told me her teams had built 47 AI agents across the organization. When I asked how many of those agents knew the other 46 existed, she laughed. "Maybe three."
This is the dirty secret of enterprise agentic AI right now. The technology works. The agents are capable. And almost nobody has figured out how to run more than a handful of them without the whole thing becoming an ungovernable mess.
I've been writing about this — in Deloitte Insights, in the Wall Street Journal — because I think the agent governance problem is the most underestimated risk in enterprise AI today. Not model accuracy. Not hallucinations. Not even security. It's the basic organizational question of: who keeps track of all these things, and how?
We've been here before
If you were doing enterprise technology in the 2000s, you watched something eerily similar happen with web services and SOA. Teams built services independently. Each had its own standards, its own security model, its own data contracts. It worked fine at small scale. Then someone needed Service A to talk to Service B, and everything fell apart.
The solution, eventually, was the service registry. A central catalog where services could be discovered, their interfaces understood, their dependencies mapped. Not sexy. Not technically innovative. But absolutely essential for operating at scale.
Agents need the same thing. Except the stakes are higher, because agents don't just expose data — they *take actions*. A misconfigured web service returns bad data. A misconfigured agent makes a bad decision and executes on it before anyone notices.
I watched this happen at a major bank in the early 2010s. They had over 200 web services built by different teams, each with its own authentication model. When regulators asked for a complete inventory of what systems could access customer data, nobody could answer. It took nine months and a dedicated team to untangle it. We're heading for the same reckoning with agents, except agents don't just read data — they act on it. The blast radius of an ungoverned agent is orders of magnitude larger than an ungoverned web service.
What an agent marketplace actually looks like
When I say "marketplace," people immediately think App Store. That's partly right, but it misses the governance layer. An enterprise agent marketplace isn't primarily about distribution. It's about four things:
Discovery. Can a team in risk management find out that a team in compliance already built an agent that does document extraction? Today, in most enterprises, the answer is no. The marketplace is the registry — it makes the organization's agent inventory visible.
Standards. Does every agent expose its capabilities in a consistent way? Can Agent A hand off work to Agent B without a custom integration? This is where protocols like MCP and A2A start to matter — but protocols alone aren't enough. You need organizational standards for how agents describe themselves, what they can do, and what permissions they require.
Governance. What autonomy level does each agent operate at? Who approved it for production? What data can it access? What actions can it take? When was it last audited? This is the part nobody wants to build because it's not flashy, but it's what keeps you out of regulatory trouble.
Observability. Once agents are in production, can you see what they're doing? Not just logs — real performance metrics. Are they making good decisions? Are they drifting? Are two agents conflicting with each other? This is where agent operations becomes a real discipline, not just a monitoring dashboard.
If I were building this for that bank with 47 agents, the first 30 days would be pure inventory and classification. Catalog every agent, who built it, what framework it uses, what data it accesses, what actions it can take, and who owns it. Days 30 to 60: establish the registration standard and make it a gate for any new agent deployment. No registration, no production access. Days 60 to 90: instrument the top ten agents by business impact with consistent observability — same metrics, same dashboards, same escalation paths. That gives you the foundation to scale. Everything before that is building on sand.
The platform play nobody's talking about
Here's what I find interesting about this problem: whoever solves agent governance for the enterprise will own the most strategic layer of the AI stack.
Think about what happened with cloud. AWS didn't win because it had the best virtual machines. It won because it built the management layer — IAM, CloudWatch, CloudTrail — that made it possible for enterprises to actually *operate* in the cloud safely. The infrastructure was necessary but not sufficient. The governance was the moat.
The same dynamic is playing out with agents. The model providers (OpenAI, Anthropic, Google) give you the reasoning engine. The framework providers (LangChain, CrewAI, Autogen) give you the scaffolding. But nobody has convincingly solved the enterprise management layer: how do you operate a fleet of hundreds or thousands of agents, built on different frameworks, accessing different systems, with consistent governance and observability?
The honest answer is that nobody has won this yet. The hyperscalers — AWS, Google, Microsoft — are all building agent management layers into their platforms, but they're optimizing for their own ecosystems. Most enterprises run multi-cloud, multi-framework environments, which means the management layer needs to sit above any single vendor. My advice: build your own registry and governance standards internally. Buy observability tooling. And wait on orchestration platforms until the market shakes out — the cost of switching will be high, and it's too early to bet on a winner.
What to do now
If you're running an AI practice at a large enterprise — and this is specifically who I'm writing for — here's what I'd do today, not next quarter:
Inventory what you have. You probably don't know how many agents are running across your organization. Find out. This is unglamorous work and it will take longer than you think.
Establish naming and registration standards before you need them. Every new agent gets registered in a central catalog with a consistent description of what it does, what it accesses, what autonomy level it operates at, and who owns it. Make this a gate for production deployment.
Pick your interoperability bet. MCP, A2A, or something else — the specific protocol matters less than the decision to standardize. Delaying this decision is the most expensive choice you can make, because every agent built without a standard is an agent you'll have to retrofit later.
Build observability into the first agent, not the fiftieth. If you can't see what your agents are doing now, you definitely won't be able to see what fifty of them are doing. Instrument early.
One thing I'd warn against specifically: don't let your agent marketplace become a bureaucratic bottleneck. I've seen organizations overcorrect from "no governance" to "governance committee approves every agent deployment in a six-week review cycle." That kills velocity and pushes teams back into shadow IT. The marketplace should be enabling, not blocking. Think of it like a well-run airport — clear rules, efficient processes, real-time monitoring — not a border crossing with a two-hour queue.
Prakul Sharma is the AI & Insights Practice Leader at a Big 4 consulting firm. He has written extensively on agent marketplaces and governance in Deloitte Insights and the Wall Street Journal. He writes weekly at prakulsharma.ai.