On 9 July, at the AI for Good Summit in Geneva, the International Telecommunication Union (ITU) unveiled an initiative trying to bring order to a field where the pace of development outstrips any attempt at control. The UN’s digital agency starts from a hard fact: AI agents are becoming autonomous so fast that the ability to trust them lags behind. The answer is a program to keep them «identifiable and accountable» — two adjectives that, when read next to «agent», stop being accessories and become the core of a deployment strategy.

The ITU’s move looks at first glance like a classic soft-law exercise: no enforcement power, no immediate technical standard. But the structural signal is different. In an ecosystem where trust is earned through transparency of decision chains, stating that agents must be identifiable and accountable amounts to putting a price on governability. And that price, for many organizations, is paid by controlling the entire stack.

Consider those who handle sensitive data on-premises, from pharmaceutical labs to central banks and critical infrastructure. An autonomous AI agent that makes decisions without the possibility of tracing why or who trained it is, today, unacceptable. Not out of ideology, but because regulators are beginning to demand audit trails and traceability of actions, especially when the agent operates beyond a purely advisory role. The ITU initiative does not introduce new obligations, but it crystallizes a market expectation: soon it will no longer be enough to trust an agent just because «it works». You will have to prove it.

The identifiability aspect is particularly relevant for those choosing on-premise. In a self-hosted environment, the agent can be given a certain digital identity (think private PKI, service certificates, hardware attestation), and every action can be recorded in immutable, verifiable logs aligned with local retention policies. In the cloud, the same transparency is often mediated by proprietary interfaces that do not expose internal mechanisms. The difference is not merely technical: it is the line between calculated trust and delegated trust.

The geopolitical context also matters. The ITU is an arena where countries with very different approaches to AI governance meet. Launching such an initiative means placing agent accountability into a multilateral framework, before individual jurisdictions do so with incompatible rules. For companies already operating with distributed architectures and local data, this is a quiet advantage: controlling the execution environment simplifies adherence to any future accountability norm, without having to retool a cloud system that responds to someone else’s logic.

There’s a shift in perspective that must be grasped immediately. Until yesterday, the question was: «Did the agent do the right thing?». Starting today, with the ITU’s imprint, the question becomes: «Can we prove why it did it?». And the answer, in many regulated sectors, is built only by starting from hardware and software that you control down to the last bit. It’s not a matter of performance or cost: it’s the difference between an agent you can put into production and one that stays confined to a lab.