AI agents are moving toward a world where they may not just complete tasks, but decide who should do them. A new paper from Google Deepmind researchers argues that this shift creates a delegation problem: current methods are not strong enough for networks where humans and AI systems pass authority, responsibility, and accountability between one another.
The proposed answer is a framework for "intelligent AI delegation." It covers how roles should be defined, how trust should be established, and how a system should decide whether a task belongs with a human, an AI agent, or another agent in the network.
What AI delegation has to solve
The researchers frame AI delegation through ideas borrowed from human organizations. One of the central concepts is the principal-agent problem. A principal gives work to an agent, but the agent's goals may not line up with the principal's goals.
In human organizations, that gap can involve incentives, oversight, or communication. With AI agents, the paper points to alignment problems such as reward hacking, where a system exploits loopholes in the objective it has been given.
Delegation also depends on how many agents a person or orchestrating system can monitor reliably. If too much is delegated without enough oversight, a supervisor may lose the ability to detect when something has gone wrong. If too little is delegated, the network may fail to capture the benefits of agentic systems.
The source also highlights the "authority gradient" concept from aviation. Large competence gaps between superiors and subordinates can suppress communication. In AI systems, a related concern is sycophancy: an agent may tell the user what they want to hear instead of challenging a task it should not accept.
Verification becomes the organizing rule
The Deepmind framework rests on five pillars. They are continuous agent evaluation, dynamic task redistribution when conditions change, traceable documentation of decisions, reputation systems for open marketplaces, and safeguards that stop individual failures from spreading through the network.
The most important principle is "contract-first decomposition." In plain terms, a task should only be delegated when its result can be checked. If the result is too subjective, too expensive to verify, or too complex to judge, the task should be split into smaller pieces until the outputs can be evaluated.
This makes verification the gatekeeper for AI delegation. The question is not only whether an agent can perform a task. It is whether the system can confirm that the task was performed correctly.
The researchers also argue for decentralized marketplaces rather than centralized directories. In their view, smart contracts could help protect both sides when tasks are distributed among agents and humans.
Security risks are part of the delegation problem
Delegating work across AI agent networks creates security concerns beyond ordinary task management. The paper flags malicious agents that steal data or return falsified results. It also points to "agentic viruses," described as self-propagating prompts.
Another risk is "cognitive monoculture." If too many agents depend on the same small set of foundation models, a single weakness could affect large parts of the network. In that kind of system, variety is not just a design preference. It becomes a safeguard against broad failure.
These risks matter because delegation creates chains of dependency. One agent's output may become another agent's input. If the first result is compromised, the error can move through the network unless checks and containment mechanisms are built in.
Why deliberate inefficiency may be useful
One recommendation stands out because it runs against the usual goal of automation. The researchers suggest that AI systems should sometimes assign tasks to people even when the system could do the work itself.
The reason is skill preservation. The idea comes from the "Automation Paradox." If AI handles every routine task, human supervisors may lose the hands-on experience they need when they have to intervene. A system can look efficient in normal operation while becoming fragile in moments of failure.
This creates an uncomfortable tradeoff. Perfect short-term efficiency may weaken long-term resilience. Keeping humans involved in some work can preserve understanding, judgment, and the ability to step in when the delegation chain breaks.
The paper also warns about a "moral crumple zone." That term describes a setup where humans sit inside a delegation chain without real control, but still absorb liability when something fails. In that case, human involvement is not meaningful oversight. It is a way to assign blame.
Existing protocols do not fully meet the bar
The researchers looked at existing agent protocols, including Anthropic's MCP, Google's A2A, and the Agent Payments Protocol. According to the paper, none fully satisfies the requirements for this kind of delegation.
The source article notes two specific gaps. MCP offers binary access without finer-grained authorization levels. A2A lacks support for cryptographic result verification.
Those limitations point back to the central argument. AI delegation is not simply a question of connecting agents to tools or to each other. It requires clear authority, verifiable outcomes, documentation, reputation, and safeguards that can hold up when tasks move across a mixed network of humans and machines.
For organizations using AI agents, the practical implication is direct: delegation should be designed around what can be checked, who remains accountable, and how human expertise is preserved. Without those pieces, automation may create systems that appear capable while becoming harder to supervise, correct, and trust.