Microsoft's 2026 Work Trend Index measured a 15-fold year-over-year increase in active agents across the Microsoft 365 ecosystem, 18-fold in large enterprises. Whatever your position on the AI cycle, that curve describes something real: software is moving from suggesting actions to taking them.
A chatbot that drafts a reply and an agent that sends it are different objects. The first has a user. The second has a principal, in the old legal sense: someone on whose behalf it acts. And principals answer for their agents. That has been true of employees and contractors for centuries. It does not stop being true when the agent is software.
The consultancies have arrived at the same point. Accenture built its Age of Co-intelligence report with Wharton on task-level analysis across 18 industries, covering more than 120 million workers, and the line Fortune pulled from that work is the one worth keeping: intelligence may be scalable, but accountability is not. You can spin up a hundred agents this afternoon. You cannot spin up a hundred people willing to sign for what those agents did.
So the practical question for any company deploying agents this year is not which framework. It is: for each agent, who is the principal? Who defined what it may do? Where is the record of what it did? Who reviews that record, and how fast can they stop it?
If those questions have no written answers, the agent does not have oversight. It has spectators.
In the EU this is where the AI Act stops being abstract. The deployer duties in Article 26 are built exactly around this: use systems according to instructions, assign competent human oversight, monitor operation. Article 12 makes logging a design requirement, which is what turns "what did the agent do" from archaeology into a lookup. Article 4, in force since February 2025, requires providers and deployers to support AI literacy among the staff operating the system, without prescribing a level. The Act has been in application since 2 August 2026, the high-risk duties under Annex III arrive on 2 December 2027, and the penalties chapter has applied since August 2025.
None of this requires slowing down agent adoption. It requires naming a human per agent and writing four answers per system. Companies that do it will scale agents the way IBM's data says controlled companies do: more, not fewer.
We draft the oversight and documentation set, article by article. klariq.eu.