The AI industry is currently obsessed with ‘agents’—autonomous programs that do more than just chat. Most current multi-agents systems are based on hard-coded, fragile heuristics, which fail to adapt when their environment changes.
Google DeepMind Researchers propose a novel solution. The research team argued that for the ‘agentic web’ to scale, agents must move beyond simple task-splitting and adopt human-like organizational principles such as authority, responsibility, and accountability.
Defining ‘Intelligent’ Delegation
In standard software, a subroutine is just ‘outsourced’. Intelligent delegation It is different. This is the sequence of actions where an authority or responsibility from a delegator to a delegates. It involves risk analysis, capability matching and trust building.
Framework of the Framework of the Framework
For this project, our research team developed 5 key requirements that were mapped onto specific technical protocol:
| Framework Pillar | The Technical Implementation | Core Function |
| Dynamic Assessment | Task Decomposition & Assignment | Granularly determining the agent’s state and capability. |
| Adaptive execution | Coordination Adaptive | Handling context switches and runtime problems. |
| Transparency of the Structure | Monitoring & Verifiable Completion | The auditing of both the processes and final outcomes. |
| Scalable Market | Trust & Reputation & Multi-objective Optimization | Efficient, trusted coordination in open markets. |
| The Systemic Resilience | Security & Permission Handling | Avoiding cascading failures, malicious use. |
Engineering Strategy: ‘Contract-First’ Decomposition
Most significant change is contract-first decomposition. According to this principle, the delegator can only assign a job if it is possible to verify its outcome.
If a task is too subjective or complex to verify—like ‘write a compelling research paper’—the system must recursively decompose it. The process continues until all sub-tasks are verified using the available tools such as formal mathematical proofs or unit tests.
Chain of Custody Recursive verification
If you are part of a chain delegating authority, for example 𝐴 → 𝐵 → 𝐶Accountability is a transitive concept.
- Agent The B-word Verification of the work is the responsibility of The C-word.
- When Agent The B-word The result is returned to You can also find out more about the A-Team here.The chain must be complete and include all cryptographic signatures.
- Agent You can also find out more about the A-Team here. Next, the check is performed in two stages: verification The B-wordDirect work of a person and that The B-word Verified Correctly The C-word.
Security Tokens and Tunnels
The security of these chains is at risk. Data Exfiltration, Backdoor Implanting” Model Extraction.
DeepMind suggests a way to protect your network. Delegation capability tokens (DCTs).. Based on technologies such Macaroons The following are some examples of how to use Biscuits, these tokens use ‘cryptographic caveats’ to enforce the principle of least privilege. An agent, for example, might be given a token allowing it to read a Google Drive folder while preventing any writing operations.
Evaluation of Current Protocols
The research team examined whether industry standards were ready to adopt this framework. While these protocols provide a base, they all have ‘missing pieces’ for high-stakes delegation.
- MCP (Model Context Protocol): Standards the way models are connected to tools. Gap in the Economy: There is no policy to control permissions in deep delegation chains.
- A2A (Agent-to-Agent): Manages the discovery lifecycle and tasks. Gap in the Economy: There are no standard headers to indicate Zero-Knowledge Proofs or digital signature chains.
- Agent Payments Protocol (AP2): Spending funds is authorized by the agent. Gap in the Economy: The quality of work cannot be verified before payment is released.
- UCP (Universal Commerce Protocol). Standardize commercial transactions Gap in the Economy: It is optimized for shopping/fulfillment, not abstract computational tasks.
The Key Takeaways
- You can move beyond heuristics. AI delegations are still based on hard-coded simple heuristics, which can be fragile. They cannot adapt dynamically to changes in environment or unanticipated failures. A framework for adaptive delegation is required to incorporate transference of authority, accountability, and responsibility.
- ‘Contract-First’ Task Decomposition: For complex goals, delegators should use a ‘contract-first’ approach, where tasks are decomposed until the sub-units match specific, automated verification capabilities, such as unit tests or formal proofs.
- Transparency in the Chains of Accountability In long delegation chains (e.g., 𝐴 → 𝐵 → 𝐶), responsibility is transitive. B is accountable for C’s work, but A also needs to confirm that B has verified C’s statements.
- Achieved security via tokens To prevent systemic breaches and the ‘confused deputy problem,’ agents should use Delegation Capability Tokens (DCTs) that provide attenuated authorization. To ensure agents adhere to the principle of minimum privilege, access is restricted to subsets and allowed operations.
Click here to find out more Paper here. Also, feel free to follow us on Twitter Don’t forget about our 100k+ ML SubReddit Subscribe Now our Newsletter. Wait! What? now you can join us on telegram as well.
