Project NANDA: Building an Open Internet for AI Agents

Ramesh Raskar, a professor at MIT and director of Project NANDA, along with core contributor Maria Gorskikh, unveiled the foundational infrastructure for what they envision as an "open internet for AI agents." This initiative, born out of Project NANDA, aims to build a decentralized architecture for networked AI agents, drawing parallels to the development of the open web for documents. The core problem NANDA seeks to solve is the current lack of shared standards for AI agents to discover each other, establish portable identities, maintain trust independent of single platforms, and transact or coordinate across organizational boundaries without permission.

The Need for an Open Agentic Web

The internet is on the cusp of hosting trillions of autonomous agents that will negotiate, delegate, and migrate at high speeds. Raskar highlighted that the existing identity and discovery systems, like DNS, were built for documents and are not equipped to handle the scale and nature of this emerging agent ecosystem. He drew an analogy to the early days of the internet, comparing the current state of AI agents within proprietary platforms to the AOL era of the late 1990s, where closed networks and gated directories limited interoperability. The goal is to transition to an "open web" for agents, enabling seamless discovery, task delegation, payment, and learning across different entities.

Project NANDA's Three Layers: Discovery, Commerce, and the Bazaar

The proposed infrastructure is built upon three key layers: Discovery, Commerce, and what Raskar termed "the Bazaar." Discovery focuses on how agents find each other, Commerce on how they transact and establish trust, and the Bazaar on how they collaborate and improve collectively. Project NANDA has already shipped components for these layers, including the NANDA Index, registries, protocols, and the NandaTown simulator.

The NANDA Index: A Discovery Layer for Agents

Maria Gorskikh provided a foundational definition of an agent as "a model that uses tools in a loop," taking a goal, choosing an action, calling a tool, and observing the result. She emphasized that for agents to perform real work, they need access to external tools and applications, underscoring the importance of open-source, self-hosted agents like Open Claw for user control. The challenge then becomes how these distributed agents can find each other. The NANDA Index addresses this by providing a shared discovery layer. Unlike DNS, which maps a name to a fixed address, the NANDA Index provides more detailed "agent cards" containing information about an agent's capabilities, rules, and communication protocols. Messages are routed through a "message box" which handles authentication, spam filtering, and queuing, ensuring that private details are not exposed. The resolution process is designed to be adaptive, changing based on the querying agent, the agent's current status, and access permissions.

Onboarding and Hosting Agents

Making agents discoverable on the NANDA Index is facilitated through platforms like host39.org. Users can sign up, fill out an "agent facts" form to generate an agent card, and publish it to the index. The project offers multiple onboarding paths, from enterprises running their own catalogs to individuals using host39.org for a hosted agent URL without needing their own domain. For hosting, agents can run locally for full control or on cloud platforms. The presentation highlighted Maritime as an agent-specific hosting platform designed to reduce costs for running multiple agents by utilizing a sleep/wake architecture, ensuring that idle agents do not consume compute resources.

NandaTown: Simulating the Agentic Web

The true challenges of an agent web lie in the interaction and coordination of thousands of agents at scale. To address this, Project NANDA developed NandaTown, an open-source discrete event simulation playbook. NandaTown models the entire agent economy, including discovery, identity, registries, messaging, and coordination. It allows developers to run and test agent interactions in a simulated environment, visualize agent networks, and analyze protocol performance. The platform supports both scripted agents and real AI models, enabling rigorous testing of coordination problems, negotiation, and fault recovery scenarios. NandaTown breaks down the agentic web into 12 protocol layers, allowing developers to test individual components or entire systems.

The Future of Agent Interaction

Project NANDA's vision is to create an open and interoperable infrastructure for the next era of the internet, where trillions of AI agents can seamlessly discover, transact, and coordinate with each other. The core components, discovery via the index, commerce through portable identity and trust, and coordination via open protocols, are being developed and tested to ensure a robust and scalable agentic web.