AI agent delegation
Delegating to AI agents · Agent-to-agent delegation
AI agent delegation is the act of entrusting a task or a decision to an AI agent, or from one agent to another, within a scope defined in advance. The moment of delegation is when a person states an intention and a system chooses and executes in their place.
From choosing to stating an intention
With traditional software the user picks an option. With an agent the user states a goal, and the agent picks the options. This looks like a small change in interface and is in fact a transfer of power, since the user has agreed to a principle and rarely to each of its consequences. The same logic applies inside companies, where a manager who asks an agent to clear a backlog has also delegated every trade-off made along the way.
Delegating in widening circles
The classic mistake is to automate too broadly and too quickly. A sounder sequence starts on a narrow scope with clear rules and active supervision, and widens only after evidence. Rollout goes by stages with a way back at each one, and the criteria for moving on are written in advance. Moving an agent from a test environment to real data counts as a new deployment: in a case reported by The San Francisco Standard in February 2026, an email agent that had behaved well on a small test inbox began deleting messages once connected to a much larger real one, and ignored the stop message sent from a phone.
When agents delegate to agents
Chains add a difficulty of their own, in that several agents that are each well governed can together produce a decision nobody authorized. No individual control fails and nobody can be blamed for the result. Governing a chain therefore takes three additions: each mandate declares what the agent depends on and what depends on it, a cumulative ceiling applies to the final decision whatever each link was allowed to do, and the chain has its own responsible person who reviews it as a whole.
Frequently asked questions
What should never be delegated to an AI agent?
See non-delegable decisions. In short: what causes irreparable harm, what touches people's physical safety, and acts the law reserves to a person.
How do you keep control when an agent delegates to another agent?
By applying limits to the outcome and not only to each link: a cumulative ceiling on the final decision, declared dependencies in each mandate, and one person responsible for the chain.
Where does your organization stand?
ACF Score measures your agentic governance in about 10 minutes, the AI Act checker tells you which obligations apply, and ACF Control applies your rules before an agent acts.
Related terms
Primary sources
Written by Vincent Dorange, creator of the Agentic Commerce Framework®. Last updated .