中文
AI Engineer World's Fair

12 minutes on the three layers of infrastructure the internet of agents is missing, and how MIT fills them with an open index and a sandbox you can actually run

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

12 min
Agent

12 min total·Actually worth watching closely: ~3 min·3 must-watch clips

Orange = the 3 minutes worth watchingFor the rest, the guide is enough
Segment guide · 7 segments
  1. 0:02 1:40Listen

    Opening vision: from a web of documents to the agentic web

    Raskar argues the internet will host eventually trillions of autonomous agents that negotiate, delegate and migrate between hosts in milliseconds, a load that strains the identity and discovery systems built for documents, DNS among them.

    The agent web isn't an extension of today's web; it needs an infrastructure of its own

    Purely spoken framing with no critical visuals. Fine to listen to like a podcast.▶ Jump to 0:02
    Speaker · Ramesh Raskar
  2. 1:40 3:30Listen

    The AOL analogy: the three things walled gardens don't give you

    Today's proprietary agent stores and orchestrations that only talk to themselves look like the AOL era of the late 90s. What's missing: cross-vendor discovery, portable identity and trust, and an open way to transact and coordinate.

    Hold on to those three gaps, because every Nanda module that follows fills one of them

    An argument-driven analogy. You get all of it by ear.▶ Jump to 1:40
    Speaker · Ramesh Raskar
  3. 3:30 5:26Skim

    The Nanda index: a discovery layer for agents

    Introduces the Nanda index as an open-source discovery layer: agents register and get an agent card, and a signed agent facts record says who the agent is, what it can do and who built it, so you can check before you connect.

    The point of a discovery layer is verification before connection, not just a directory listing

    Mostly conceptual slides. A glance at the structure diagram while listening is enough.▶ Jump to 3:30
  4. 5:26 7:04Watch

    Inside the index: the message box and adaptive resolution

    Messages don't reach the agent's runtime directly. A message box checks who's sending, handles access, filters spam and bad requests, and holds the message until the agent is ready. And resolution isn't one fixed name-to-address mapping: it changes based on where the agent is, who is asking, and what they're allowed to access.

    Adaptive resolution is the real break from DNS: one agent can have many endpoints and traffic can be routed accordingly

    The mechanism diagram at 326s (email-style identity, messages flowing through the message box) is the key visual for the whole design. Worth pausing on.▶ Jump to 5:26
    Speaker · Maria
  5. 7:05 9:05Skim

    Getting your agent onto the index: three low-friction paths

    One path per kind of user: individuals fill out the agent facts form at host39.org and publish an agent card; enterprises run their own catalog and register a gateway from their own domain; existing websites can later use DNS AID to connect agents to domains they already own, and small businesses get a hosted agent URL.

    The bar for getting listed is very low, and you can go live without owning a domain

    Step-by-step slide content. Skim it and remember the three paths.▶ Jump to 7:05
    Speaker · Maria
  6. 9:05 10:20Listen

    Hosting and cost: keeping an agent online

    An agent is only useful if it stays reachable. A self-hosted gateway (open-source OpenClaw) gives you full control, cloud hosting fits most cases, and at scale Maritime's sleep-and-wake architecture keeps idle agents from burning compute.

    The hard part of scale isn't getting agents running, it's not paying for them while they idle

    The comparison is spoken. Just catch which hosting option fits which case.▶ Jump to 9:05
    Speaker · Maria
  7. 10:20 12:10Watch

    Nanda Town: proving the protocols in simulation

    The open-source sandbox breaks the agent web into 12 parts (transport, communication, identity, registry, auth, trust, payments, coordination, negotiation, memory, privacy, data facts) and runs marketplace negotiation, auctions, voting and supply chain experiments as a discrete event simulation to see where things break.

    You don't assume the protocols hold up, you find where they break in simulation before they're load-bearing on the real internet

    The most visual stretch of the talk: the live sandbox with messages moving on a map at 620s, and the 12-part architecture overview at 678s. Both need the picture.▶ Jump to 10:20
    Speaker · Maria