From Web Pages to AI Agents: The Paradigm Shift
The internet was built for documents. AI agents need something fundamentally different — and the NANDA protocol is building it.
The Web We Built vs. The Web We Need
The World Wide Web was a triumph of simplicity. Tim Berners-Lee's original vision — hyperlinked documents served over HTTP — scaled from a physics lab at CERN to billions of pages in under two decades. DNS mapped human-readable names to IP addresses, TLS certificates authenticated servers, and the browser became humanity's universal interface.
But in 2026, the web is undergoing a transformation as fundamental as the shift from dial-up to broadband. The clients are no longer just humans with browsers — they are autonomous AI agents that discover peers, negotiate tasks, delegate sub-goals, and maintain persistent memory across sessions. As MIT's Ramesh Raskar and the Project NANDA team argue in their foundational paper Upgrade or Switch, the question is no longer whether the web's infrastructure needs to change — it's how radically.
Three Eras of Web Architecture
To understand why agents break the web, trace the evolution through three architectural eras:
Era 1 — Static Web (1990s)
HTML files, CSS, images. Client fetches a URL, server returns bytes, conversation ends. Stateless, immutable, human-initiated.
Era 2 — APIs & Services (2000s–2010s)
REST, GraphQL, serverless functions. Machines became first-class clients — but still reactive, stateless, and endpoint-bound.
Era 3 — Autonomous Agents (2024+)
Goal-directed, persistent, proactive. Agents pursue objectives, maintain memory, delegate sub-tasks, and spawn helper agents — without constant human direction.
The leap from Era 2 to Era 3 is not incremental. As the Upgrade or Switch research demonstrates, agents introduce three new thresholds that no existing web infrastructure was designed to handle: self-directed discovery at millisecond latency, delegated authority with instant revocation, and cryptographic proof of behavior.
The Dial-Up Lesson
The NANDA research team draws a clarifying analogy: the transition from dial-up to broadband. When the internet was first commercialized, existing telephone infrastructure seemed like a natural fit — it already connected most homes. But dial-up's 56 kbps ceiling and circuit-switched architecture couldn't support the always-on, bandwidth-intensive applications that the internet would demand.
Engineers didn't try to make phone lines faster forever. They designed purpose-built last-mile infrastructure — DSL, cable, fiber — that accommodated known requirements while leaving room for unknown future needs. The open TCP/IP substrate they built allowed applications like video streaming, social media, and cloud computing to flourish in ways no one predicted.
What Agents Actually Need
The Upgrade or Switch paper identifies five categories where existing web infrastructure falls short for autonomous agents:
- Sub-second discovery — DNS propagation takes minutes to hours. Agents spawning helper agents need discovery in milliseconds.
- Capability-based trust — TLS certificates prove domain ownership, not agent capabilities. An agent delegating a financial task needs to verify that the recipient can actually execute trades, holds proper credentials, and has a clean behavioral history.
- Privacy-preserving lookups — DNS queries expose access patterns to network observers. When a healthcare agent searches for a specialist, that search pattern itself is sensitive.
- Dynamic routing — Agents migrate between runtimes, scale across regions, and update capabilities continuously. Static DNS records can't keep up.
- Instant revocation — CRL/OCSP mechanisms can't handle real-time trust revocation at trillion-agent scale. When a delegated agent misbehaves, its privileges must be revoked in milliseconds, not hours.
NANDA: Purpose-Built Infrastructure
This is the problem that Project NANDA (Networked AI Agents in Decentralized Architecture) was created to solve. Rather than stretching 1983-era DNS or bolting agent metadata onto WHOIS records, NANDA provides a purpose-built discovery, trust, and routing layer for the agentic web.
The architecture centers on three components:
- The NANDA Index — a decentralized agent registry using the "Quilt" federation model, where multiple registries operate independently while maintaining global interoperability. Records are ultra-lean (≤120 bytes) for efficient global-scale indexing.
- AgentFacts — W3C Verifiable Credentials that serve as cryptographically signed "resumes" for agents, containing capabilities, endpoints, trust scores, and compliance attestations.
- Protocol Bridges — cross-platform interoperability connecting A2A, MCP, and HTTPS endpoints so agents built on any protocol can discover and work with agents on any other.
As Forbes described it: "They're making TCP/IP for AI, and it's called NANDA."
The Road Ahead
The transition to an Internet of AI Agents won't happen overnight. Like the move from dial-up to broadband, it will likely follow a hybrid path — existing DNS infrastructure for human-facing services, purpose-built NANDA infrastructure for agent-to-agent interactions, with bridge protocols enabling interoperability between the two.
Project NANDA's three-phase roadmap reflects this reality: Phase 1 builds the foundational discovery and identity layer (the NANDA Index), Phase 2 introduces economic primitives for agentic commerce, and Phase 3 enables Large Population Models and cross-silo collaborative learning.
What's clear is that the paradigm has already shifted. The web is no longer just for documents and APIs — it's becoming a substrate for autonomous intelligence. The question isn't whether we need new infrastructure for this reality, but whether we'll build it deliberately or discover its absence the hard way.