Project Nanda is building open infrastructure for agents to discover, identify, communicate, and transact across organizational boundaries.
2
The NANDA index gives agents portable identity, signed facts, adaptive routing, and a shared way to find other agents.
3
Nanda Town tests discovery, trust, payments, negotiation, and coordination in simulated agent networks before they run at internet scale.
Summary
Ramesh Raskar presents Project Nanda as an open infrastructure project for a web of AI agents. He compares today's agent ecosystem to the AOL era, where closed platforms and proprietary directories limit communication between systems. An open agent web would let agents from different organizations discover each other, hand off work, pay for services, and learn across boundaries. The talk explains the NANDA index, which publishes agent identities, capabilities, access rules, and endpoints through signed agent facts and agent cards. It also covers hosting options and Nanda Town, an open-source discrete-event simulation for testing agent networks. Nanda Town models discovery, messaging, identity, trust, payments, negotiation, memory, privacy, and other parts of an agent economy. The practical message is that connecting one agent is easy, while testing trust and coordination among many independent agents requires shared infrastructure and simulation.
The agent web needs open infrastructure beyond today's closed platforms
Ramesh Raskar says current agent development often happens inside walled gardens, proprietary agent stores, and orchestration systems that only communicate with themselves. He compares this period to the AOL era, when users lived inside a closed network and gated directory. The open web changed that model by allowing people to create websites and browsers that could communicate without one company's permission. Raskar expects a similar transition for agents. An agent from one organization should be able to discover another agent, hand off work, pay it, and learn from it across organizational boundaries. He frames the system around discovery, commerce, and the bazaar.
Maria defines an agent as a model that uses tools in a loop: it receives a goal, chooses an action, calls a tool, reads the result, and continues until the task is done. Once agents run locally, on clouds, and on servers controlled by different people, they need a way to find each other. The NANDA index provides that shared discovery layer. It publishes who an agent is, what it can do, and how to reach it. Unlike DNS, which maps a name to an address, the index gives agents information about capabilities, tools, rules, and communication methods.
Agent cards and signed facts let agents check who they are contacting
An identity such as agent@hotmail.com is resolved by the NANDA index into an agent card. The card describes the agent, explains how to reach it, and identifies where messages should go. Messages first arrive at a message box, which checks the sender, handles access, filters spam or bad requests, and holds messages until the agent is ready. A signed agent facts record gives other agents information about the agent's identity, capabilities, permissions, builder, and endpoint. This lets an agent check basic facts before connecting.
The index can route requests without exposing one fixed endpoint
The NANDA index is not a simple name-to-address table. It can return updated agent facts based on the request, and one agent can have multiple endpoints. Resolution can change according to where the agent is, who is asking, and what that requester is allowed to access. This allows traffic to be sent to a suitable endpoint while keeping private details hidden. Raskar's proposed onboarding paths include enterprise catalogs, DNS AID for existing websites, and host39 for individuals or small businesses that need a hosted agent URL without owning a domain.
Hosting agents requires different choices for control and cost
An agent can run locally, which gives its owner full control but requires the owner to keep it online. Cloud hosting is more practical for many use cases. Raskar names AWS as a general cloud option and Maritime as an agent hosting platform intended to make hosting cheaper and simpler. Maritime uses a sleep-and-wake architecture so idle agents do not keep consuming compute. The problem becomes more significant when a team, product, or simulation needs many agents running at once, because the cost of each agent affects whether the system is practical.
Nanda Town tests agent networks before they carry real workloads
Nanda Town is an open-source simulation project from Project Nanda. It models an agent economy with discovery, identity, registries, messaging, and coordination. Users can view agents on a map, watch messages move, compare protocol results, and replay runs step by step. Existing experiments include buyer and seller negotiations, auctions, voting, consensus, and supply chains. These scenarios test how agents make deals, reach decisions, pass messages, and recover from failures. Nanda Town is small enough to run on a laptop.
The simulator treats the agent web as a set of replaceable layers
Nanda Town divides the agent web into 12 parts: transport, communication, identity, registry, authentication, trust, payments, coordination, negotiation, memory, privacy, and data effects. The NANDA index is only the registry component within this larger system. Developers can replace one layer with their own implementation and run it alongside the rest of the network. Scenarios are defined in a short YAML file, then agents and traffic are injected into a discrete-event simulation. Tier one uses scripted agents, while tier two uses real AI models.
"You can take one layer, add your own version, run it inside Nanda Town, and see how it works with the rest of the network."11:22
Who should watch
You are building an agent that needs to communicate with agents owned or hosted by other organizations.
Your system needs portable identity, capability records, message filtering, or adaptive routing instead of a single platform's directory.
You want to test trust, payments, negotiation, or coordination across many agents before deploying the system.
