941: Multi-Agent Human Societies — with Dr. Vijoy Pandey
Dr. Vijoy Pandey of Cisco's Outshift incubator discusses his vision for a future where multi-agent human societies collaborate to solve major scientific and physical challenges. He introduces AGNTCY, an open-source project for the "Internet of Agents," designed to address the critical hurdles of agent interoperability and trust through a Zero Trust framework and Task-Based Access Control (T-BACK).
A Vision for Multi-Agent Human Societies
In the future, societies will be composed of both humans and AI agents, collaborating to solve humanity’s most significant problems, from understanding the universe to discovering new medicines and materials. This vision extends beyond the “myopic and boring” application of agents in business software. It encompasses agents for scientific discovery, social interaction, and, crucially, embodied agents (physical AI or robotics) that can handle monotonous physical tasks like washing dishes, freeing humans for more creative and enjoyable pursuits.
This collaborative future requires a platform where diverse agents—language models, scientific foundation models (like AlphaFold), embodied robotics, and compliance agents—can all interoperate seamlessly. This need for an open, heterogeneous platform, not siloed to a single vendor or cloud, is the driving force behind the “Internet of Agents” concept.
AGNTCY and the Internet of Agents
The “Internet of Agents” is a thesis for an open and interoperable framework, much like the original internet, that allows agents and humans to collaborate effectively. AGNTCY (spelled A-G-N-T-C-Y) is the open-source manifestation of this thesis. Hosted by the Linux Foundation, its formative members include Cisco, Dell, Google, Red Hat, and Oracle.
The project aims to provide the necessary infrastructure for the AI era, enabling collaboration between agents themselves and between agents and humans. It acknowledges the need for multiple communication protocols and aims to provide a stable scaffold around them, addressing key functions:
- Discovery: A “DNS for agents” to find agents and their capabilities.
- Identity: A system to manage who or what an agent is and what it represents.
- Communication: The protocol layer for interaction (collaborating with projects like Google’s A2A and MCP).
- Evaluation: A framework to measure agent performance and ensure they are working towards a common goal.
Blending Deterministic and Probabilistic AI
A critical insight is the need to blend the deterministic tools of traditional computing with the probabilistic nature of modern AI. The current excitement around LLMs has led many to ignore the power of deterministic systems like knowledge graphs and ontologies. The real magic happens in the combination of these two worlds, a concept known as Neuro-symbolic AI. This approach leverages the strengths of both, achieving speed and accuracy simultaneously. A practical example is using a deterministic digital twin of a network in combination with probabilistic generative AI agents to validate network configurations efficiently and reliably.
The Trust Crisis: Identity and Access Management (IAM) for Agents
Traditional IAM systems, built for humans and deterministic software, break down when applied to AI agents. The reasons are twofold:
- Probabilistic Nature: Agents are not predictable. Their behavior, intent, and outcomes can change from one moment to the next. You can ask an LLM the same question five times and get five different answers.
- Machine Speed and Scale: Agents operate at a speed and scale that humans cannot. Granting an agent broad, persistent permissions, as is common in Role-Based Access Control (RBAC), creates an enormous security risk. An agent could act on behalf of one user one moment and another the next, all at machine speed.
The Solution: Zero Trust for Agents (ZTA) and T-BACK
To solve this, we must return to the first principles of security, leading to a framework called Zero Trust for Agents (ZTA). The core of this framework is a new access control model: Task-Tool-Transaction Based Access Control (T-BACK or T3-BACK).
Instead of trusting an agent based on a persistent role (e.g., “Vijoy’s agent”), T-BACK operates on a “zero trust” principle:
- Access is granted for a single, specific, and granular task, tool usage, or transaction.
- Privileges are elevated just-in-time for only the duration of that specific task.
- Permissions are immediately revoked the moment the task is complete.
For this to work, three key infrastructure components are required:
- Semantic Parsing: An engine that can analyze the discourse between agents to understand the specific task being requested.
- T-BACK in Identity Providers: Authorization servers must support the ability to issue these short-lived, task-specific tokens.
- Ephemeral Runtimes: Sandboxed environments where an agent is allowed to execute its approved task with elevated privileges, and is then immediately removed.
How to Get Started with AGNTCY
For developers and data scientists interested in building more secure and interoperable agentic systems, the AGNTCY project is accessible and open for contribution.
- The main portal is agency.org, which provides documentation and links to the GitHub repository.
- A key resource is “Coffee Agency,” an open-source reference application. It models a complex coffee supply chain, providing a real-world example where developers can plug in their own components (e.g., different databases, messaging buses, identity providers) to see how the AGNTCY ecosystem works in a practical setting.
The Next Frontier: Semantic Interoperability
The current work on the Internet of Agents focuses on syntactic interoperability—getting frameworks like LangChain, CrewAI, and various vendor SDKs to communicate, akin to standardizing grammar. The next, more profound challenge is semantic interoperability. This involves ensuring that agents share a common understanding of the meaning behind their communication. When one agent says a solution is “best” or has “90% confidence,” does that mean the same thing to another agent from a different system? Solving this requires tackling the complex domains of knowledge representation and cognition, moving us closer to the vision of truly collaborative multi-agent human societies.