Google Cloud sets four rules for AI agent delegation

Google Cloud has outlined four principles for how AI agents should delegate work in multi-agent systems, drawing on research from Google DeepMind.

The framework is aimed at organisations building AI agents that can divide tasks, assign work to other agents and coordinate activity across business workflows. It focuses on verification, cost control, data access limits and safeguards against automatic compliance.

Verification first

The first principle is that delegated work should be designed to be checked. Google Cloud describes this as a contract-first approach, in which an orchestrating agent breaks a larger assignment into smaller tasks that can be monitored and verified.

That may involve repeatedly splitting a job into sub-goals until each step is simple enough to assess reliably. Where that is not possible, organisations may need human judgement to determine whether an outcome meets expectations.

The aim is not to remove people from the process, but to identify where human oversight is most useful. In practice, that means reserving expert review for parts of a workflow that cannot be graded or validated automatically.

Cost choices

The second principle focuses on cost. Google Cloud says customers are increasingly asking whether specific tasks can be assigned to smaller, cheaper models instead of larger reasoning systems.

Under the framework, an effective delegating agent should match each task to the model or endpoint best suited to the job. A complex business process such as payroll may require a more advanced model, while a basic administrative task such as spreadsheet reformatting may not.

This has become more prominent as companies try to balance AI spending against reliability. Some customers use model routing through API gateways to direct tasks to different models, while others rely on client-side proxy tools.

Data boundaries

The third principle deals with sensitive information. Orchestrating agents should pass on only the minimum permissions and data a sub-agent needs to complete a task.

That supports both security and efficiency. Sending a full set of payroll or other confidential data to a secondary agent could expose information unnecessarily while also increasing the amount of context the system must process.

A harder question is how one agent can prove work was completed correctly without disclosing the underlying private data. Google Cloud points to advanced cryptographic methods, including zero-knowledge proofs, as one potential answer.

In that model, an agent could show that a computation was performed properly without revealing the dataset used in the process. That would allow the delegating agent to verify the outcome while preserving confidentiality.

Compliance risk

The fourth principle concerns what organisational theory calls the zone of indifference, where a worker accepts a task without challenge because it appears to fall within normal duties. Google Cloud argues that a similar pattern can create risks in agent-based systems.

Many current AI systems are shaped by post-training safety filters and system instructions, and often comply with requests unless a clear violation is triggered. In a chain of delegated requests, that tendency could allow errors, mismatched intent or harmful instructions to pass through several agents without scrutiny.

The research addressed that risk directly.

"As delegation chains lengthen (? → ? → ?), a broad zone of indifference allows subtle intent mismatches or context-dependent harms to propagate rapidly downstream, with each agent acting as an unthinking router rather than a responsible actor."

Google Cloud says the answer is what the research calls dynamic cognitive friction. Agents should test whether a request is accurate, relevant and sufficiently clear, and should be able to challenge ambiguous instructions or seek human verification when needed.

Human oversight also needs to be used carefully. Too much intervention could burden staff and reduce the efficiency gains automated systems are meant to bring, while too little could leave risky decisions unchecked.

The broader argument is that delegation should be treated as an intelligence problem rather than a simple hand-off mechanism. Instead of merely routing tasks onward, multi-agent systems need to decide what can be verified, which model is suitable, what information can be shared and when a request should be questioned.

Google Cloud says its longer-term aim is to make agents work more naturally within organisations, including in systems where tasks are delegated between software agents and human specialists. Intelligent delegation, it says, will need to respect organisational boundaries and involve human experts only where necessary.