Who signs the agent's mandate?
Every company has a policy that says who can sign an invoice for 50,000 zloty. Almost none has a policy that says what an agent deploying changes to production a hundred times an hour is allowed to do.
Two poles of the same revolution
In recent weeks two stories have appeared that are worth putting side by side.
The first is Spawn, a ten-person startup building an engine for multiplayer games. Its founder Jacob Sansbury described on X a company in which permanent agents with names, memory and real permissions are responsible for the engine, finance, bugs and deployments. According to his account, the company spends $3.1 million a month on tokens and pushed 77,000 changes in 30 days. A change goes to production before anyone reviews it. An hour without a deployment is a top-severity incident.
More interesting than the numbers, though, is what happened when 700 items were waiting in the queue for Sansbury’s decision. The founder built an “AI Judge”, an agent that decides on his behalf based on two years of his previous decisions. Within eight days the Judge handled more than 11,000 cases: it resolved 7,296 itself, sent 2,213 back for rework and passed 1,863 to humans. Sansbury personally decided 27 times.
The second story is Anthropic, which in October published a report on how AI has changed work inside the company itself. There, the human stays in the loop. The result? Agent working hours exceed human working hours six to one. According to the report, about 80% of the time in an interactive session is spent waiting for a human to respond. Median time spent in meetings rose 20% year on year. One engineer describes his work as facing an army of Claudes, each asking for attention.
And one more number from that report that should interest every board: code written by an agent was no more often deleted or reverted later than code the same person wrote by hand in the same month.
So quality is ceasing to be the bottleneck. The decision is becoming the bottleneck.
The wrong question
The discussion about agents in companies usually revolves around whether the human should be in the loop or out of it. Spawn answers “out”, Anthropic answers “in”. In my view both answers are incomplete, because the question is wrongly put.
The right question is: who gives the agent its mandate, within what scope, and who can take that mandate away.
This is not a technology question. It is a corporate governance question. Anyone who has run a company or sat on a supervisory board knows the delegation of authority matrix. It is clear who approves the budget, who signs a contract above the threshold, who can hire and who can fire. The matrix exists so that the organisation can move fast where the risk is small and slowly where a mistake is expensive.
Agents need exactly the same. Meanwhile, in most companies I see in my work with clients and fund portfolio companies, an agent’s permissions follow from whichever systems it happened to get API keys for. That is not a mandate. That is an accident.
Three classes of decisions
The simplest way to bring order to this is to divide decisions by the cost of getting them wrong.
Reversible, cheap decisions. Fixing a bug, updating data in the CRM, matching an invoice to an order. Here the agent should act on its own, and the human only sees a trace in the log. Keeping a human in the loop for such decisions is exactly the 80% of time that agents at Anthropic spend waiting.
Reversible but costly decisions. Changing the price list, halting a shipment, replying to a key customer. Here the agent prepares the decision and its rationale, and a human approves it. It is worth agreeing in advance what happens if the human does not respond within a set time, because otherwise the agent stands still and the company waits.
Irreversible decisions. Deleting data, signing a commitment, a statement to a regulator. Here the human decides, and the agent can at most prepare the material. Spawn, for all its radicalism, puts hard gates in exactly this place.
This division is not a discovery. What is striking is how rarely anyone actually writes it down.
Questions the board should ask management
If I sat today on the supervisory board of a company deploying agents, I would ask three questions.
First: who signs the agent’s mandate? Is there a specific person accountable for the fact that the finance agent may approve an invoice correction but may not change payment terms?
Second: who can revoke that mandate, and how quickly? Spawn describes a situation in which, after an API key leak, the agents themselves detected the problem and rotated all keys within 15 minutes. That is impressive. But the board needs to know who will stop the agent when the problem is the agent itself.
Third: how do we measure whether the mandate is set correctly? The share of cases escalated to humans says more than the number of agents deployed. If there are too many escalations, the agent is just a more expensive form. If there are too few, someone has stopped looking.
How we approach this at Sophona
When building AI Operators at Sophona, we start with the mandate, not the model. Before an Operator touches its first process, we agree with the client which decisions it makes on its own, which it prepares for approval and where its role ends. Context memory exists so that the Operator knows not only what to do, but also when to wait for a human.
I don’t think every company should become Spawn. Most shouldn’t. But every company that lets agents into its processes should be able to answer the question Spawn answered for itself directly: where is a mistake irreversible, and who decides then.
Technology can be fixed in the next release. A mandate that nobody wrote down cannot be fixed after the fact. It is worth writing it down before an agent does it for us.

