Videos sum9DgexFRQ
The Agentic Web and the Bazaar Era of AI - Ramesh Raskar, MIT Media Lab
Scene timeline
28 shot(s).
keyframes kept every frame deduplicated
What was stored
- cues
- 132
- whisperx 132
- chunks
- 22
- from 132 cues
- keyframes
- 27
- kept of 28 captured
- frames with text
- 26
- 527 lines read
- chapters
- 0
- from the source metadata
- keyframe bytes
- 2.2 MB
- word timings on 132 cues
Provenance
| stage | state | model | started | took |
|---|---|---|---|---|
fetch |
done | — | 2026-08-11 03:24 | 1m 17s |
stt |
done | — | 2026-08-11 03:25 | 13s |
chunk |
done | — | 2026-08-11 03:26 | 0s |
text_embed |
done | — | 2026-08-11 03:26 | 0s |
keyframe |
done | — | 2026-08-11 03:26 | 35s |
ocr |
done | — | 2026-08-11 03:26 | 14s |
frame_embed |
done | — | 2026-08-11 03:26 | 5s |
Frames, and what the machine read
-
- 01 /260.91
- PROJECT NANDA0.99
- AN OPEN AGENTIC WEB0.96
- An open stack for a web of agents0.99
- The infrastructure being built so agents discover, hand0.99
- off to, pay, and learn from eachother, across0.98
- organizational boundaries, no permission required.0.98
- projectnanda.org1.00
-
- SPEAKERS1.00
- Ramesh Raskar1.00
- Maria Gorskikh0.97
- Professor, MIT0.97
- Core contributor, Project NANDA1.00
- Director, Project NANDA0.98
- projectnanda.org1.00
-
- Project1.00
- NANDA1.00
- The interet of Al Agents0.93
- NETWORKED AI AGENTS IN A DECENTRALIZED ARCHITECTURE0.99
- WHY IT'S NEEDED0.97
- An open layer the agent web doesn't have yet0.99
- 01 Agents must find each other across vendors,0.99
- and no shared directory exists today.1.00
- 02 They need portable identity and trust0.99
- Open research building the infrastructure for an Internet of0.99
- not owned by any single platform.0.99
- Al Agents the way the open web was built for documents.0.99
- 03 They need to transact and coordinate across0.98
- organizations, with no permission required.0.99
- NANDA Index0.99
- Registries1.00
- Protocols1.00
- NandaTown1.00
- projectnanda.org1.00
-
- THE PREMISE0.99
- billions→trillions0.99
- of autonomous agents on the internet, that negotiate, delegate,1.00
- and migrate in milliseconds.1.00
- DNs-centred identity & discoverywon't carry that load. The agent web needs its own.0.99
- projectnanda.org1.00
-
- ONE COMPANY'S GARDEN0.98
- THE PATTERN0.99
- We've been here before1.00
- Build an agent today and you build it0.98
- inside a walled garden, closed0.99
- platforms, proprietary agent stores,1.00
- orchestration that only talks to itself.0.99
- This is the AOL era of Al. What came1.00
- after AOL was the open web.1.00
- gated1.00
- locked1.00
- no1.00
- directory1.00
- runtime1.00
- exits1.00
- projectnanda.org1.00
-
- THE TRANSITION0.99
- Closed1.00
- d→Open→projectnanda.org1.00
- The next era needs what the web needed: open0.98
- infrastructure, discover across vendors, hand off work,1.00
- settle payment, learn together. It comes in three layers.0.99
- 031.00
- Bazaar1.00
- agents get better together0.98
- 021.00
- Commerce1.00
- agents pay each other1.00
- 011.00
- Discovery1.00
- agents find each other1.00
- projectnanda.org1.00
-
- LEVEL-SET1.00
- What is an agent?0.99
- A model using tools in a loop.0.97
- 011.00
- 020.98
- 031.00
- 041.00
- Goal1.00
- Choose action0.99
- Call tool0.99
- Observe result0.98
- repeat until done, memory, orchestration & multi-agent systems build on top of this loop1.00
- projectnanda.org1.00
-
- EVEL-SET1.00
- What is an agent?0.99
- A model using tools in a loop.0.98
- 011.00
- 021.00
- 031.00
- 041.00
- Goal1.00
- Choose action1.00
- Call tool0.94
- Observe result0.99
- repeat until done, memory, orchestration & multi-agent systems build on top of this loop1.00
- projectnanda.org1.00
-
- BUILD THE LOOP0.97
- Open frameworks1.00
- CLOSED AGENT1.00
- Locked to one vendor's runtime0.99
- Anchor on the open-source ones,1.00
- OpenClaw is a good example: an open1.00
- framework for building and running agents.1.00
- OPEN AGENT1.00
- You can't build an open agentic1.00
- Portable by design1.00
- web on closed agents, that just0.99
- moved· inspected· hosted anywhere0.96
- rebuilds the walled garden one0.99
- level down.0.98
- projectnanda.org1.00
-
- BUILD THE LOOP0.97
- Open frameworks1.00
- CLOSED AGENT1.00
- Locked to one vendor's runtime1.00
- Anchor on the open-source ones,1.00
- OpenClaw is a good example: an open1.00
- framework for building and running agents.1.00
- OPEN AGENT1.00
- You can't build an open agentic1.00
- Portable by design1.00
- web on closed agents, that just1.00
- moved· inspected · hosted anywhere0.96
- rebuilds the walled garden one0.99
- level down.0.98
- projectnanda.org1.00
-
- THE DISCOVERY LAYER1.00
- The NANDA Index0.99
- If the protocols are the roads of the0.99
- agent city, the Index is the directory of0.99
- THE INTERNET TODAY1.00
- who lives where.1.00
- DNS → name to address1.00
- It's the discovery layer.1.00
- But agents need more than an address.1.00
- projectnanda.org1.00
-
- ESOLUTION PATH0.99
- low the Index works0.94
- Identity0.97
- NANDA Index0.99
- Agent Card1.00
- MessageBox1.00
- Runtime1.00
- agent:john@...0.99
- resolve → entry0.97
- facts + inbox URL0.98
- auth · queue0.96
- private1.00
- "id": "urn:ai:email:john",0.98
- Global discovery, local control, nobody owns the whole thing.1.00
- "messagebox": {0.99
- A portable, privacy-preserving identity even for someone who doesn't own a domain.0.99
- "url": "mbox/john",0.97
- The runtime can move providers without changing the public identity, and it's never the0.99
- "auth": ["oauth"]0.96
- public endpoint.1.00
- "runtimeEndpoint": "not_public"0.99
- nn0.54
- projectnanda.org1.00
-
- identity1.00
- who the agent is, portable,0.99
- vendor-free1.00
- ERIFIABLE ATTESTATION1.00
- AgentFacts1.00
- capabilities what it can do, credentialed0.98
- assertions1.00
- A portable, cryptographically signed1.00
- record, not an address, but a verifiable0.99
- permissions1.00
- what it's allowed to touch1.00
- claim about what an agent is and can do.1.00
- provenance1.00
- who built it & who vouches for it0.99
- signed1.00
- time-to-live1.00
- tamper-evident1.00
- endpoints1.00
- where to reach it, one or many0.98
- projectnanda.org1.00
-
- NOT A STATIC LOOKUP0.99
- Resolution is adaptive1.00
- 00.56
- A lean index resolves identity to dynamic, verifiable AgentFacts, the directory stays0.99
- thin; the work happens at query time.1.00
- 011.00
- 021.00
- 031.00
- Multi-endpoint routing1.00
- Load balancing1.00
- Privacy-preserving access1.00
- one agent, answerable from several places0.99
- spread traffic across those endpoints0.99
- prove a capability without exposing0.98
- everything1.00
- projectnanda.org1.00
-
- host39.org1.00
- HIPPED, NOT SLIDEWARE0.98
- H391.00
- host391.00
- nost39 is live1.00
- Sign up, fill the AgentFacts form, and get a0.99
- AGENT CARD HOSTING1.00
- hosted agent URL today, listed on the1.00
- host your agents1.00
- NANDA Index.0.99
- at predictable URLs0.99
- host an0.98
- host39.org1.00
- agent1.00
- host39 is a third-party A2A agent card host. Register your identity, create agent0.98
- cards, and publish them at stable public URLs — no server required.0.99
- the index1.00
- nandaindex.org1.00
- Get started free1.00
- Go to dashboard0.95
- projectnanda.org1.00
-
- NBOARDING1.00
- Four ways onto the Index0.99
- 011.00
- 021.00
- 031.00
- 041.00
- Enterprise1.00
- DNS-AID1.00
- SMBs1.00
- Personal1.00
- Al Catalog0.97
- Agents1.00
- Resolve agents through the1.00
- Sign up at host39.org and fill1.00
- Clone & deploy the nanda-1.00
- DNS you already own, meeting1.00
- the AgentFacts form to get a1.00
- The same host39 path, a1.00
- registry, then register your1.00
- the existing web where it is.0.99
- hosted URL, no domain0.99
- portable, vendor-free agent1.00
- gateway in the Index from your1.00
- required.1.00
- identity for an individual.1.00
- own domain.0.97
- • Live0.95
- Planned1.00
- • In testing0.92
- •In testing0.97
- projectnanda.org1.00
-
- HOSTING1.00
- Where do agents live?1.00
- 01 Local0.94
- 021.00
- Cloud1.00
- Full control. Great for development. Your data never1.00
- General infra like AWS, or agent-specific platforms like0.98
- leaves the machine.0.98
- Maritime for persistent, shareable hosting.1.00
- keeping it up & reachable is on you0.99
- convenience over control1.00
- Hosting a single agent is a solved problem. But not every cloud costs the same, and that0.99
- matters more than you'd think.0.96
- projectnanda.org1.00
-
- maritime1.00
- MONTHLY COST1.00
- ALWAYS-ON AGENT1.00
- HE CLOUD DEFAULT1.00
- Cheap enough to run a0.97
- maritime1.00
- $1/mo flat0.99
- fleet of agents1.00
- AWS Lambda1.00
- $18+/mo1.00
- AWS EC20.95
- $8-15/mo1.00
- Maritime hosts an agent for a flat $1/month.1.00
- Fly.io1.00
- $6+/mo1.00
- Sleep/wake means it only burns compute1.00
- Railway1.00
- $5-10/mo1.00
- while it's actually thinking, idle agents cost1.00
- almost nothing.1.00
- ~1 vCPU / 1 GB, always-on equivalent.0.98
- Real deployments aren't one agent,0.99
- The flat rate comes from sleep/wake, no0.99
- charge while idle.1.00
- they're dozens or hundreds. At that scale,1.00
- cost-per-agent is the budget.0.98
- projectnanda.org1.00
-
- THE HARD PART1.00
- The hard part isn't running one agent1.00
- It's everything that happens between agents.1.00
- 011.00
- 021.00
- 031.00
- 041.00
- Discover1.00
- Identify0.99
- Trust1.00
- Coordinate1.00
- find each other1.00
- prove who they are1.00
- decide who to rely on1.00
- work as a group1.00
- One agent is easy. Thousands of them interacting is the real test, so you run it and watch0.99
- what breaks.1.00
- projectnanda.org1.00
-
- PROVING IT OUT0.99
- Infrastructure for the agentic era1.00
- So how do you prove it works? You simulate it.1.00
- projectnanda.org1.00
-
- SANDBOX TOWN1.00
- NandaTown1.00
- A simulation playbook for the infrastructure1.00
- 121.00
- of the agentic web. The whole agent1.00
- economy, discovery, identity, registries,1.00
- protocol layers1.00
- messaging, coordination, modeled so you can1.00
- run it and watch what breaks first.0.98
- 61.00
- 21.00
- Don't argue about how an agent web0.99
- scenarios1.00
- agent tiers1.00
- should work. Spin one up and watch how1.00
- it behaves.0.98
- Open source1.00
- Apache 2.00.97
- Project NANDA1.00
- projectnanda.org1.00
-
- INSIDE THE SANDBOX0.98
- It's a live, running town0.99
- ·Streaming0.95
- github.com/pro1.00
- nandatown.projectnanda.org/agents1.00
- US-CA1.00
- 110.91
- Cambridge Project NANDA0.98
- 341.00
- US-MA1.00
- Bengaluru1.00
- 310.95
- IN1.00
- São Paulo0.99
- New York0.96
- 271.00
- US-NY1.00
- Tokyo1.00
- 241.00
- JP0.94
- London1.00
- 221.00
- UK1.00
- Agents, the live0.96
- Live feed, who's talking0.99
- Visualizer, replay a run, message by1.00
- Leaderboard, how1.00
- network map1.00
- to whom1.00
- message1.00
- protocols rank1.00
- projectnanda.org1.00
-
- WHAT WE'VE BEEN RUNNING1.00
- Set up a1.00
- Experiments, already live1.00
- agents, get0.96
- nandatown.projectnanda.org/experiments1.00
- II0.67
- MARKETPLACE1.00
- 100 ADENTS0.85
- AUCTTON0.95
- VOTING0.97
- 22 AGENTS0.92
- CONSENSUS1.00
- T AGENTE0.87
- SUPOLY CHAIN0.96
- AGENTO0.98
- Marketplace: 100 Agents0.98
- Auction: 50 Bidders1.00
- Voting: Proposal & Election0.99
- Quorum Consensus: 7 Nodes1.00
- Supply Chain: 4-Hop Pipeline1.00
- 50 buyers and 50 sellers negotiate prices.0.98
- An auctioneer posts items; 49 bidders0.98
- A proposer submits topics, 20 voters cast1.00
- A leader proposes values; followers vote to0.98
- Goods move from supller to manufacturer0.98
- Buyers make offers; sellers accept or reject.0.99
- compete. The highest bid wins each round.0.98
- yes/no ballots, and a coordinator counts the1.00
- commit or abort. The value is committed0.99
- to distributor to retailer. Each hop processes0.99
- Tests how well agents find fair prices witho.0.98
- Tests whether agents settle on fair prices...0.97
- results. Tests simple group decision-making.1.00
- when enough votes agree. Tests leader-..0.99
- and passes ther on. Tests whether the..0.99
- 100%1.00
- DELIVERY0.89
- 2.lt0.79
- LATENCE0.83
- 2,2001.00
- 100%1.00
- L8t0.98
- LATENDY0.77
- 1,8501.00
- 100%1.00
- BELIVERY0.81
- 1.2t0.85
- LATEUDY0.78
- 9801.00
- 100%0.90
- DELOVERY0.89
- It0.93
- LATENCY0.96
- UGOE0.57
- 4201.00
- 100%1.00
- DELTVERY0.65
- 440.83
- 3401.00
- View detalis↓0.95
- Visualize →0.90
- View detalls ↓0.94
- Visualize →0.84
- View detalls0.94
- Visualze →0.82
- View details↓0.96
- Visualize →0.90
- View detalls↓0.93
- Visualize→0.88
- projectnanda.org1.00
Transcript
132 cues· 1,893 words· 10,327 chars
- 0:01 Welcome, everyone.
- 0:03 Over the next few minutes, we're going to talk about the open infrastructure being built for a web of agents, why it's needed, what's already shipped, and exactly where your own agent fits into it.
- 0:17 It comes out of Project Nanda, an open research effort that started at MIT.
- 0:23 And by the end of this presentation, you will know how to put an agent that you build yourself on the open web all by yourself.
- 0:31 So let's get into it.
- 0:36 I'm Ramesh Raskar, professor at MIT and director of Project Nanda.
- 0:41 And Maria is a core contributor to Project Nanda.
- 0:46 So first, what is NANDA and why does it need to exist?
- 0:50 NANDA, which stands for Networked AI Agents in a Decentralized Architecture, is an open research building the infrastructure for an internet of AI agents, the way open web was built for documents.
- 1:05 The gap it fills is concrete.
- 1:07 Agents have no shared way to find each other across vendors, no portable identity or trust that isn't owned by a single platform, and no open way to transact and coordinate across organizations.
- 1:20 Nanda builds that missing layer and ships it in open.
- 1:24 The index, the registries, the protocol, and the Nanda Town simulator you will see later.
- 1:34 Here's the premise that drives all of it.
- 1:37 The internet is about to host not millions or billions, but eventually trillions of autonomous agents.
- 1:44 They negotiate, they delegate, they migrate between hosts in milliseconds.
- 1:49 That's a fundamentally different load than the human web.
- 1:52 And it strains the identity and discovery system we've built for documents, DNS among them.
- 2:01 The web that's coming needs infrastructure of its own.
- 2:09 And we have been here before.
- 2:10 If you're building agents today, you're mostly building them or you're forced to build them inside walled gardens, closed platforms, proprietary agent stores, and orchestrations that only talk to itself.
- 2:23 And it kind of works, but it also feels familiar because we have been here before.
- 2:29 This is like the AOL era from the late 90s, where it was a closed network.
- 2:34 You know, AOL, you got the CDs and you installed it on your PC.
- 2:38 It was a closed network.
- 2:39 It was a gated directory.
- 2:41 You live inside the garden created by this one company called AOL.
- 2:46 And what came after AOL was this open web.
- 2:49 Everybody's in a permissionless manner, creating websites, creating browsers, and any website can talk to any browser.
- 2:57 And that's the transition that's about to happen to agents as well.
- 3:03 So the next era needs what the web needed, an open infrastructure, where an agent from one company or one entity can discover agent from another.
- 3:13 this agent can hand off work to it, pay it, learn from it across organizational boundaries, no permissions required.
- 3:20 There are three layers here, discovery, commerce, and what I will call the bazaar.
- 3:27 So hold on to those three concepts.
- 3:29 We'll come back to them and, and discussed.
- 3:31 So first the basics.
- 3:35 Hey, my name is Maria and I'm a core contributor to Project Non-Done.
- 3:39 So let's start with a simple definition of an agent.
- 3:42 The way I think about it, an agent is a model that uses tools in a loop, right?
- 3:47 You give it a goal, it decides what to do next, it calls a tool, it looks at the result, then it keeps going until the task is done.
- 3:55 So that loop is the core idea.
- 3:57 Everything else like memory, orchestration, and multi-agent systems is built on top of it.
- 4:04 So you can build this agent loop in many different ways.
- 4:07 One example is Open Claw.
- 4:09 So Open Claw is a self-hosted agent gateway.
- 4:12 That means you can run it yourself and connect it to the apps and tools you already use.
- 4:18 This matters because agents are not just chatbots.
- 4:20 If an agent is going to do real work, it needs to access your real tools and apps.
- 4:27 And if it has access to real tools, we should care about who controls it, where it runs, and how much we can see.
- 4:35 And that is why open source self-hosted agents are super important.
loading