Nine out of ten AI agents can be well-governed. The tenth can bring the whole thing down. That's the question MWC 2026 raised and couldn’t answer.
By Jana Sedivy, VP Customer Experience & Product, InkBridge Networks
I joined Drew Clark on Broadband Breakfast Live from the floor of Mobile World Congress 2026 in Barcelona, alongside executives from Mimosa Networks, Ookla, and Dentons. We covered a lot of ground - AI in telecom, digital sovereignty, satellite convergence, and a surprising turn on supply chain risk (I've written up the full set of themes in my main MWC 2026 wrap-up).
The nine-good agents problem
Todd Daubert, a partner at Dentons with more than two decades of MWC attendance, raised a point about agentic AI that the industry has been sidestepping.
Paraphrasing: you can have a whole series of AI agents talking to each other. Nine of them can be well-governed, well-behaved, and solidly engineered. The tenth can bring the whole thing down.
That is the agentic AI problem for telecom in one sentence. And as of MWC 2026, the industry doesn't have a good answer for it - not in regulation, not in engineering practice, not in procurement.
What agentic AI looks like in telecom
The agentic AI being deployed in telecom networks is autonomous software that can take actions across systems - rebalancing network loads, spinning up capacity, reconfiguring radio parameters, investigating anomalies, opening tickets, closing tickets, calling other agents to help.
A single agent operating inside well-defined boundaries is manageable. A mesh of agents working together at machine speed is where the governance question gets sticky. The efficiency case for agentic AI depends on agents acting without asking permission at every step. That same property is the risk.
For anyone running network infrastructure, this is a familiar problem in a new shape. We've spent decades working out how to make sure the right device, user, or service gets on a network and does only what it's permitted to do. Agentic AI is asking the industry to solve that problem all over again for software entities that can spawn other software entities.
The absence of accountability
When a human operator in a network causes harm - intentionally or through negligence - there's a chain of accountability. Employment records, access logs, contracts, professional licensing in some cases. We've built 50 years of network operations practice around the assumption that a human is attached to every action.
When an AI agent causes harm, that chain breaks.
Case in point: a recent incident in the open-source community provides a concrete example: an autonomous AI agent, rejected by a project maintainer, went on to write and publish a targeted hit piece about him and tag him in the project's discussion thread. No human instructed it to do any of that. The human operator behind the agent eventually came forward, called it a "social experiment," and faced no consequences. I’ve written elsewhere about that case and what it says about AI agent governance in more depth in a separate post.
The telecom version of that problem is bigger.
If an autonomous agent operating in a carrier's network misbehaves - takes an action it shouldn't, leaks data it shouldn't, loops into a cascade with other agents that brings down a service - who answers for it?
Today, the answer is usually "we'll figure it out after the fact," which is another way of saying nobody has figured it out in advance.
What regulation 2026 should address
Todd's point on the panel was that regulation should create a stable platform for the benefits of technology rather than prescribe the technology itself. I agree with that framing. The question is what that platform needs to include for agentic AI specifically.
Three things would go a long way:
- Persistent AI agent identity. Every autonomous agent operating in production infrastructure should have a traceable identifier tied back to the deploying organisation. A liability regime, so to speak. You know hwo your access points are. You should know who your agents are.
- Action logging standards. Agents operating across organisational boundaries (and in telecom, they will) need common audit trail requirements so forensic investigation is possible after the fact.
- Clear liability attribution. The human or organisation that deployed the agent should bear the consequences of its actions by default, until regulation or contract says otherwise.
None of this is technically difficult. The current obstacle is that the industry's marketing momentum is running in the opposite direction. Agentic AI is being sold on the promise of hands-off autonomy, and accountability frameworks slow that story down. The alternative is a governance vacuum, and vacuums tend to get filled by whoever moves first, usually badly.
Watch the full panel
The full Broadband Breakfast Live panel from MWC 2026 is worth watching - Todd's point on agentic AI governance was one of several moments that deserved more airtime than we had.
If you're thinking about where autonomous AI agents fit into your network operations and how to build authentication and access controls that account them, get in touch. We've been solving the "who gets to do what on this network" problem for 25 years. The answers change when the "who" isn't human, but the discipline is the same.
Related Articles
Mobile World Congress 2026: what ISPs and network operators should take away
AI dominated the conversation at Mobile World Congress 2026, but the more interesting story was about who controls the infrastructure it runs on.
Open source is paying the price for AI agent governance failures
An AI agent had its code rejected by an open-source maintainer, then autonomously published a hit piece on him. The governance frameworks that might have prevented this barely exist yet.