AI Engineer World's Fair 2026

The Agentic Web and the Bazaar Era of AI - Ramesh Raskar, MIT Media Lab

Loading slides…

Read the talk

The Agentic Web and the Bazaar Era of AI

Project NANDA connects independently operated agents through discovery, signed identity records, and open coordination protocols, with NandaTown as the testbed.

From a talk by Ramesh Raskar and Maria Gorskikh

How does an agent find another agent?

You can build an agent yourself. How do you put it on the open web so agents operated by other people can discover it and work with it? That is the starting problem for Project NANDA, an open research effort originating at MIT and presented here by Ramesh Raskar and Maria Gorskikh.

NANDA stands for Networked AI Agents in a Decentralized Architecture. Its goal is to provide infrastructure for agents comparable to what the open web provides for documents. The missing pieces are shared discovery across vendors, portable identity and trust outside any single platform, and open mechanisms for transactions and coordination across organizations. NANDA presents its index, registries, protocols, and NandaTown simulator as components of that open layer.

Slide titled “An open layer the agent web doesn't have yet,” with three reasons it is needed and labels for NANDA Index, Registries, Protocols, and NandaTown.Open full source frame
NANDA’s proposed open layer for agent discovery, identity, trust, and coordination.
0:010:17
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:01 · section reference included

From gated directories to an open agent web

Raskar’s architectural premise is a future internet with trillions of autonomous agents, including agents that negotiate, delegate, and migrate between hosts in milliseconds. These are projections motivating the design. Such a network would put different demands on identity and discovery than a web organized around documents and relatively stable destinations.

Today’s proprietary agent stores and isolated orchestrators resemble the AOL era: a useful network whose directory and interactions remain inside one company’s garden. The transition to the open web let people independently create websites and browsers that could work together. The corresponding transition for agents would let an agent discover another organization’s agent, hand off work, pay it, and learn from it without requiring membership in the same platform.

Raskar organizes that transition around three ideas: discovery, commerce, and the bazaar. Discovery makes independent agents reachable; commerce supports exchanges between them; the bazaar is the broader environment in which independently operated agents interact and coordinate.

1:341:49
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:34 · section reference included

The loop inside each agent

Gorskikh starts with a minimal definition: an agent is a model using tools in a loop. Given a goal, it chooses an action, calls a tool, observes the result, and repeats until the task is complete. Memory, orchestration, and multi-agent systems build on that loop.

Four boxes connected by arrows show Goal, Choose action, Call tool, and Observe result, with a note to repeat until done.Open full source frame
The agent loop: goal, choose action, call tool, observe result.

OpenClaw is her example of a self-hosted agent gateway: something you can run yourself and connect to the applications and tools you already use. Once an agent can act through real tools, control over its runtime matters. Who operates it? Where does it run? How much of its activity can you inspect?

Self-hosting gives owners more control, but distributes agents across laptops, clouds, and servers belonging to different people. Control over one agent does not provide discovery of other agents. That shared discovery problem is the job of the index.

3:353:47
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:35 · section reference included

From identity to a protected message box

The NANDA Index gives agents a shared place to publish their identity, capabilities, and contact information. DNS provides the familiar analogy: resolve a name so a client knows where to go. Agent discovery needs additional information about what an agent does, which tools it can use, which rules it follows, and how to communicate with it.

The resolution flow in the presentation starts with an email-shaped agent identity:

  1. Submit the identity to the NANDA Index.
  2. Receive an agent card describing the agent, its reachability, and where messages should go.
  3. Send a message to the message box, rather than directly to the runtime.
  4. Let the message box check the sender, enforce access, filter spam or bad requests, and hold the message until the agent is ready.

The message box separates receiving a request from having a runtime ready to process it.

An AgentFacts record supplies signed metadata about identity, capabilities, permitted access, the agent’s builder, and reachability. Another agent can inspect these facts before connecting. The signature provides a basis for checking the record; deciding whether to trust the agent still requires evaluating what those facts mean for the intended interaction.

Resolution is adaptive rather than a permanent mapping from one name to one address. An agent can have multiple endpoints, and the index can return updated AgentFacts appropriate to the request. Routing can depend on the agent’s location, the requester, and that requester’s access rights. This also allows private details to remain undisclosed when the requester does not need or cannot access them.

5:005:16
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:00 · section reference included

Publishing an agent for discovery

Gorskikh’s onboarding path starts at host39.org: complete the AgentFacts form, obtain an agent card, and publish it to the NANDA Index. Once listed, the agent becomes discoverable along with the information needed to reach it.

The entry point depends on who operates the agent:

  • Enterprises: Run their own catalog and register a gateway from their own domain.
  • Existing websites: Later connect agents to domains they already own through DNS-AID. This remains a proposed integration; DNS-AID is an individual Internet-Draft, not an established IETF standard.
  • Individuals and small businesses: Use Host39’s AgentFacts form to obtain a hosted agent URL without owning a domain.

These paths aim to make participation possible for both an organization with existing infrastructure and a person publishing a personal agent.

7:047:17
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:04 · section reference included

Keeping the runtime available

Listing an agent does not keep it running. After publishing its discovery information, the operator still needs a runtime that stays available and reachable. Gorskikh presents three hosting choices:

Hosting choiceOperational emphasis
Local machineDirect control; owner maintains uptime
General cloud, such as AWSEnterprise-oriented cloud hosting
Agent platform, such as MaritimeSimpler hosting focused on agents

The choice concerns runtime operation, separate from the index’s discovery role.

The cost concern becomes more significant when a team, product, or simulation runs many agents. Maritime’s mechanism is a sleep-and-wake architecture for OpenClaw and other agents: idle agents sleep instead of continuously consuming compute. Gorskikh presents this as a way to make larger fleets practical and inexpensive, without supplying a numerical cost comparison.

7:568:06
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:56 · section reference included

Testing the interactions in NandaTown

Getting one agent online leaves the harder questions between agents: how thousands discover one another, prove their identities, decide whom to trust, and coordinate without a central authority. Protocols need to be exercised under load, with failures observed, before applications depend on them across the public internet.

NandaTown provides an open-source simulation environment for that work. It models discovery, identity, registries, messaging, and coordination together. The interface lets you see agents on a map, watch messages move, compare protocol results, and replay a run step by step. Gorskikh describes the simulator as small enough to run on a laptop.

Its experiments make coordination concrete:

ExperimentInteraction under study
MarketplaceBuyers and sellers negotiate prices
AuctionAgents bid on items
VotingAgents submit and count ballots
ConsensusAgents reach agreement
Supply chainAgents coordinate across a chain

Across these scenarios, the questions concern how agents make deals, agree on decisions, pass messages, and recover when something breaks.

Slide titled “Experiments, already live” showing five illustrated experiment cards in the NandaTown interface, with the presenter inset at upper right.Open full source frame
NandaTown’s experiments cover marketplaces, auctions, voting, consensus, and supply chains.
9:029:13
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:02 · section reference included

Replace one layer and test the network

NandaTown divides the infrastructure into twelve parts: transport, communication, identity, registry, OAuth, trust, payments, coordination, negotiation, memory, privacy, and data facts. The NANDA Index occupies the registry role in this architecture. The current repository calls the fifth layer Auth and defaults to an in-memory registry, so the architectural roles do not imply that OAuth and a live NANDA Index are wired into every simulation.

A contributor can replace one layer and observe how it interacts with the rest of the network. The testbed runs a discrete-event simulation: define a scenario in a short YAML file, inject agents and traffic, and let the scenario unfold so its end-to-end behavior can be measured. This makes the unit of experimentation a complete interaction, even when the implementation change is confined to one layer.

The two simulation tiers separate scripted behavior from model-driven behavior:

TierAgent behavior
Tier oneSimple scripted agents
Tier twoReal AI models

Current documentation distinguishes deterministic scripted tests from nondeterministic model runs. Its default transport has zero simulated latency, so simulated durations do not establish real network performance. The tiers let protocol testing begin with controlled behavior before introducing model variability.

The resulting architecture connects discovery through the index, commerce through portable identity and trust, and coordination through open protocols, with NandaTown providing a place to test their interactions. Participation can begin with publishing one agent or implementing one infrastructure layer; it does not require building the entire agent web. The presentation closes by directing builders to projectnanda.org for papers and the latest projects.

10:5511:11
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:55 · section reference included

Resources

From the talk

Transcript download
Full transcript

Loading transcript…