In this presentation, Ramesh Raskar and Maria introduce Project Nanda, an open research initiative from MIT focused on building the infrastructure for a decentralized internet of AI agents. Unlike the current web designed for documents, the emerging web will host trillions of autonomous agents that negotiate, delegate, and migrate rapidly across hosts. Project Nanda aims to create an open, interoperable ecosystem where agents from different vendors can discover, trust, and transact with each other without being confined to closed platforms or proprietary systems. This new infrastructure is essential to support the scale and complexity of agent interactions, much like how the open web replaced closed networks like AOL in the 1990s.
Maria explains the core concept of an AI agent as a model that operates in a loop: it receives a goal, decides on actions, uses tools, evaluates results, and continues until the task is complete. Agents are not just chatbots but need access to real tools and applications, which raises concerns about control, privacy, and transparency. To address this, Project Nanda emphasizes open-source, self-hosted agents that users can run themselves, giving them full control over their agents. However, with agents distributed across various locations and owners, a key challenge is enabling them to find and communicate with each other efficiently.
The Nanda index serves as the discovery layer for the agentic web, functioning similarly to DNS but tailored for agents. It provides a shared registry where agents publish their identities, capabilities, communication protocols, and access rules. This index returns adaptive agent cards that help other agents verify who they are interacting with and how to reach them securely. Users and organizations can onboard their agents onto the index through different methods, from hosting their own catalogs to using hosted services like host39.org, making the system accessible to individuals and enterprises alike.
Running agents continuously and affordably is another challenge addressed by Project Nanda through platforms like Maritime, which offers cloud hosting optimized for AI agents with features like sleep and wake modes to reduce compute costs. However, the real complexity lies in how thousands of agents interact at scale—discovering each other, establishing trust, coordinating actions, and handling failures without centralized control. To test and refine these protocols, Project Nanda developed Nanda Town, an open-source simulation environment that models the entire agent ecosystem, allowing researchers to run experiments on marketplaces, auctions, voting, and consensus mechanisms in a controlled setting.
In summary, Project Nanda is pioneering the open infrastructure needed for a new era of AI agents on the internet. It provides the foundational layers of discovery, identity, trust, commerce, and coordination through open protocols and registries, all rigorously tested in simulation environments like Nanda Town. This work aims to enable a permissionless, interoperable agent web where agents from diverse sources can seamlessly interact, creating a bazaar-like ecosystem of autonomous AI services. Interested individuals can learn more and participate by visiting projectnanda.org and exploring their open-source tools and research.
