Applied AI
What did your agents do last month?
What did these agents do last month? Claire Knight, director of engineering at n8n, a platform for automating workflows, asked herself that question on stage at NVIDIA GTC in Berlin. It is the question her CEO could put to her. Her answer was that right now she probably could not say, and that her company and the industry are working on it.
Most organizations are in that position. According to a figure that Knight put together from several studies, 68 percent of organizations cannot say which actions their AI agents carry out.
The question sounds simple. This blog sets out what has to be in place before an organization can answer it. It draws on Knight's session and on three other sessions at the conference, which we went through afterward.
An agent acts, and the organization answers for it
A chatbot gives an answer, and a person decides what to do with it. An agent does the work itself.
In a session on the governance of autonomous AI, Sadeepa Wijesekara of Hewlett Packard Enterprise took the example of a system in a factory that fails now and then. A chatbot lists five possible causes, and a person works out which one applies. An agent that has to get the system running again checks each cause itself. It logs in to other systems, reads their log files, and may change a setting.
That takes more access than the chatbot had. Knight cited a figure from earlier this year in which 74 percent of companies say their agents hold more access than they need.
The organization answers for what happens with that access. Wijesekara expects auditors to ask why an agent took a decision, which data it used, and which version of that data. In a session on running agents in production, Michael Balint and William Benton of NVIDIA recalled the chatbot of Air Canada, which wrongly told a customer that a fare could be refunded. The customer took legal action, and the airline was held responsible for what its chatbot had said.
The difficult case is the good employee
Knight's example was Lena, an engineer at a company that works with cameras on vehicles. Her model has to recognize pedestrians, and at night it finds only 60 percent of the people in front of the vehicle. To improve it, Lena needs labeled night footage. That footage is restricted, because it contains personal data. Even the route a vehicle drives can identify a person.
Through her agent, Lena asks to work on the footage on her laptop over the weekend, then for a few clips, then for the metadata alone, the data about the clips without the video. A good engineer would make every one of these requests. All three are refused, because each would take personal data to a place where it may not go.
Nobody in this example means harm, and that makes it difficult. Existing controls notice a copy to an unexpected place or an unusual connection. They have far more trouble with a colleague who is doing the job, in an approved system, with data that belongs in that system. Organizations used to rely on human judgment there. In Knight's view, agents need more control at exactly that point.
Refusing everything by default is no solution, because then nobody can work. Rules at the level of the network see that data moves. What the data means and why someone needs it stays out of their view.
Every request passes one gateway
In Knight's demonstration, every request of an agent passes one gateway, a single point where the rules are kept. The rules are set once and apply to every agent that comes through.
The rules decide. They are deterministic, which means that the same request always gets the same answer, and whatever the rules do not allow is refused. Balint and Benton recommended the same. Before an agent acts, a check in ordinary code has the final say, and the model cannot talk its way past it.
A language model then puts the verdict into words, so that Lena reads an explanation where a system would otherwise return an error code. Knight's slide stated that the model cannot change the verdict. A small model is enough for that work.
A refusal comes with an alternative. Lena is offered a protected environment inside the company, where the training runs on the footage while the footage stays in its place. The rules are met, and Lena can do her work.
The whole setup, agent, gateway, model and data, ran side by side on one small machine in Germany, the kind of machine from our previous blog. Knight called it a demonstration of the concept and a starting point, and said the same can be built with other tools.
The log says which rule applied and why
Every request leaves a line in a log, allowed or refused. One line on Knight's slide shows what the log holds.
- the decision and the action the agent asked for, in this case starting a training run
- who asked, with name and role
- which data it concerned, how that data is classified, and in which country it has to stay
- where the work was sent
- which alternative was offered
- which model worded the explanation, and that it ran on the organization's own machine
The log also names the rule behind each decision. Knight's point was that the answer then goes beyond whether something was allowed, to which rule applied and why.
In an earlier blog, we described three requirements for AI in a public service. One of them was that the organization can prove what ran, where, and under which conditions. For agents, a log like this is that proof.
The log is kept outside the agent
A log is reliable when the agent itself cannot change it. In a session on agents in high-stakes environments, Ofir Arkin of NVIDIA pointed to recent incidents in which agents went off course in order to complete their task. An agent that goes off course can also falsify its own log.
That is why, according to Arkin, the control has to operate outside the agent, out of its reach and without its knowledge. The agent is treated as untrusted, and what it does is verified continuously.
Knight's demonstration follows the same line. The agent has one way out, and that way leads through the gateway. An attempt to copy the data around the gateway is stopped at the network. The log is kept in an environment that the gateway can reach and the agent cannot, and it can only be added to.
An agent has to be stoppable
Seeing what an agent did is one half of the answer. The other half is being able to stop it. Wijesekara asked for one button that switches all agents off at once.
Amanda Grady of ServiceNow, who spoke in the session with Arkin, started from the observation that visibility alone is not yet governance. Her answer is a passport for agents. It is a digital document that states who owns the agent, what its purpose is, which tools it may use, and how high its risk is. The passport is checked at every action. Without a valid passport, the answer is no.
In the demonstration, withdrawing the passport stopped the agent at once. ServiceNow wants to extend this, so that one permission can be withdrawn while the agent keeps working. The passport is a proof of concept, built together with NVIDIA, and ServiceNow is publishing the specification openly.
Grady's example of why this matters was the employee who builds an agent and later leaves the organization. That agent has to be retired as well.
What has no answer yet
The log shows the actions of an agent, one at a time. What the agent was trying to achieve stays out of view. Someone in the audience raised this, and Knight agreed that it is a gap.
The task of an agent also changes over time, and rules that are too precise cannot follow. Knight's answer was to write rules on the identity of an agent, on what this agent may read and where it may write.
Documents and loose text are harder to judge than data in a fixed structure. For those cases, the gateway can pass the question to a person, which in Knight's view fits the period in which the rules are still being tuned. A model that decides whether a request is allowed would, she said, probably fail a compliance review today.
Three questions before the agents start
Knight gave her audience three questions to take back to their own organization.
- Which of our agents can reach data that a person would need clearance for? Access to personnel files and customer data is limited to the people who need it. The same holds for agents.
- What did these agents do last month? The answer is a log that names the rule behind each decision.
- When did we last break one of these controls on purpose? Either the control holds, or the organization learns where it fails. Knight called this one scary and very revealing.
For a client, these questions belong in the assignment, before the first agent gets access. A supplier can show the answers. That means the log of a test run with the rule behind each decision, an agent that is stopped on request, and a control that is broken on purpose. Balint and Benton place tracing and limits in the architecture from the first day.
In an AI strategy, this is part of governance, next to data, people and suppliers. Quilyx advises on AI strategy and leads digital transformation, for enterprises and government. We work for the people who have to decide on AI and are held to account for it later.
Our principle is that humans decide and systems prepare. Every output of an AI system lands in a human workflow, and every decision stays with a person. That principle is in every design we make and in every system we run ourselves. For agents, it asks for rules that people set in advance, a log that shows what the system did under those rules, and a person who can stop it.
With that in place, the question about last month has an answer.
About this article: This article is our own reading of four sessions at NVIDIA GTC Berlin in October 2026, Bring Your Own Agent, The Economics and Governance of Autonomous AI, Delivering Agentic AI for High-Stakes Environments Through Zero-Trust Infrastructure and AgentOps 101. The ideas from those sessions belong to the speakers and are credited to them. The translation to clients, the positions and the choices in this article are ours. We wrote and edited it together with AI, which helped with the research, the working out and the wording. Quilyx carries the editorial responsibility.