AI agents should not only answer tasks but execute them independently: retrieve data, control systems, plan intermediate steps, pass on results.This very autonomy is now coming into the focus of supervisory authorities.According to current reports, both the EU and US authorities are calling for real-time monitoring of the activities of AI agents.The problem: the common agent frameworks – i.e. the software toolkits used to build such agents – do not provide the data required for this.
For companies, this creates a gap between aspiration and technology.Anyone who integrates an agent into a business process often cannot subsequently prove without gaps which decision was made on what basis, which systems the agent addressed and at which point it deviated from the expected route.
Why log data looks different for agents
With classic software, traceability is comparatively simple: input, processing logic, output.With an agent, in between lies a chain of model calls, tool accesses and intermediate decisions that can differ from run to run.A log that only records the final result is not sufficient here.
An agent run only becomes traceable once several levels are recorded:
- which task was set and which context was available to the agent,
- which tools and interfaces it called, with which parameters,
- which intermediate results led to which next step,
- where cancellations, repetitions or queries took place,
- which person or which system triggered and approved the run.
If one of these levels is missing, then in case of doubt neither an error can be analysed nor an audit request answered.
An experiment as illustrative material
How confusing the behaviour of several collaborating agents can become is shown by an experiment that is also currently being reported on: AI agents were supposed to build a civilisation together in Minecraft – and got lost in bureaucracy.Instead of advancing the actual task, the agents occupied themselves with themselves.
The experiment is not evidence of behaviour in productive enterprise systems, and the game environment is not comparable with a business process.It is still useful as an illustration: several agents that commission and coordinate with one another generate a momentum of their own.Without logs that make this coordination visible, nobody notices that effort arises without any result arising.
Consequences for project planning
The obvious reaction is to retrofit monitoring later.That is difficult with agents because the relevant information arises at the moment of execution and cannot be reconstructed afterwards.What was not recorded does not exist.
In practice, this means writing observability into the architecture as a requirement, not into the operational phase:
- Define logging before the first agent runs. Which events are captured, in which format, how long retained?
- Introduce a separate logging layer. If the framework does not provide the data, the application must generate it itself at the call points – ideally independently of the framework used, so that a change remains possible.
- Limit permissions. The fewer systems an agent can reach, the more manageable the space that must be monitored.
- Build in approval points. Steps with external impact – payments, contract documents, customer communication – should require human confirmation.
- Consider data protection too. Detailed logs often contain personal data; purpose, access and deletion periods must be defined.
What remains open
The available reports name neither concrete deadlines nor technical minimum requirements for the required real-time monitoring.Which evidence suffices in an individual case and how granular the logging must be cannot be derived from them.Companies should therefore observe developments rather than rely on a particular interpretation.
The direction is reliable: expectations of traceability are rising, while the tools in use are still lagging behind.No vendor will close this gap for all parties in the short term.
Classification for web and software projects
For projects that integrate agents into portals, editorial systems or internal processes, logging becomes a functional component with its own effort and its own budget – not a by-product of operations.Anyone who plans logging, approval points and limited permissions from the outset remains able to act in the face of future requirements and can even recognise misconduct in the first place.
A narrowly defined first use case is therefore sensible, allowing traceability to be tested in practice before agents are deployed more widely.