R2 Labs
Agent Delegation Contract
A portable specification for defining what an AI agent is authorized to decide, do, delegate, and escalate.
The Problem
Identity tells you who an agent is. Tool permissions tell you what an agent can technically call. Neither necessarily represents the complete delegated authority chain behind a task, or who is accountable when the agent acts.
As organizations move from AI that generates content to AI that takes action, that gap becomes a governance and security problem. An agent with valid credentials and a working API key can still take an action no one authorized. Existing identity and access-management models were built for people and services with relatively stable, coarse-grained permissions, not for agents that negotiate, delegate to other agents, and operate across shifting contexts.
The missing object is a delegation contract: a portable, machine-readable record of what a specific agent, in a specific context, has actually been authorized to do, and what it must prove about having done it. That record needs to travel with the task and define authority, prohibitions, tools and resources, delegation boundaries, escalation conditions, time and resource limits, evidence requirements, and failure behavior.
What ADC Defines
Six questions, answered per agent.
What an agent may decide
What an agent may do
What an agent may access
What an agent may delegate to another agent
When an agent must escalate to a human
What an agent must be able to prove about its actions
What ADC Is
- An open specification
- Runtime-neutral
- Model-neutral
- Vendor-neutral
- Designed for heterogeneous agent environments
- Focused on delegated authority and accountability
What ADC Is Not
- An orchestration platform
- An agent framework
- An identity provider
- A proprietary R2 product
- A workflow engine
- iDragonFly
Enterprise Use Cases
Where delegated authority matters.
A procurement agent
Authorized to negotiate within a defined budget and vendor list, but required to escalate any exception to a human buyer.
A security-response agent
Authorized to isolate a compromised host automatically, but required to prove the action was taken and why, and to escalate before taking irreversible remediation steps.
A multi-agent research workflow
Where one agent may delegate a subtask to another agent, and the delegation itself, not just the original instruction, needs to be traceable.
Regulated approval chains
A financial or healthcare workflow where an agent may prepare a decision, but a licensed human must approve it, and the chain of custody must be provable.
Cross-platform agent delegation
An agent built on one vendor platform delegates a task to an agent on a different platform, and the receiving agent needs a portable, verifiable record of what it is authorized to do.
Disconnected or degraded environments
An agent operating with intermittent connectivity still needs to enforce its authorized scope locally and reconcile evidence once connectivity resumes.
Adoption Path
Six steps to evaluate ADC.
Review the specification
Read the working specification and evaluate it against a real workflow.
Map an existing workflow
Identify a workflow where an agent already acts, or is about to.
Define the authority chain
Document what the agent may decide, do, access, delegate, and escalate.
Test a reference contract
Draft a contract for the mapped workflow and test it against real scenarios.
Integrate with existing systems
Connect the contract to identity, policy, orchestration, and audit systems.
Evaluate evidence and failure behavior
Confirm the system can prove what happened and fails predictably when it should.
Current Status
ADC is a working specification under active development by R2 Labs. It is early-stage and evolving. This page does not represent a claim of broad industry adoption or finalized standardization.
Review the specification on GitHub →Related Reading
The thesis behind ADC (why prompts are not enough, and why agent governance requires contracts) is developed further on MichaelERuiz.com.
Read: Why AI Agents Need Contracts, Not Just Prompts →Review the specification.
If you are designing agentic systems and need a portable way to define and prove delegated authority, we would welcome the conversation.