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

Project Nanda builds open infrastructure for a decentralized internet of AI agents, analogous to how the open web was built for documents, enabling agents to...

By Sean Weldon

Open Infrastructure for the Agentic Web: A Technical Analysis of Project Nanda's Decentralized Agent Coordination Framework

Abstract

Project Nanda addresses the fundamental infrastructure deficit in autonomous AI agent ecosystems by establishing open, decentralized protocols for cross-organizational agent discovery, identity verification, and coordination. Current agent deployments operate within proprietary platforms that lack standardized mechanisms for inter-platform discovery, portable trust frameworks, and open transaction protocols. This analysis examines NANDA (Network AI Agents in a Decentralized Architecture), a three-layer infrastructure system comprising discovery mechanisms through the Nanda Index, commerce protocols via cryptographically-signed agent facts records, and coordination frameworks validated through Nanda Town discrete event simulation. The system introduces email-like agent identity formats, adaptive resolution protocols, and sleep/wake computational architectures to enable internet-scale agent operations. By providing foundational protocols for agents to discover, authenticate, and transact across organizational boundaries without central authority, this infrastructure establishes technical prerequisites for an agentic web capable of supporting trillions of autonomous entities operating at machine timescales.

1. Introduction

The architectural requirements of the emerging agentic web diverge fundamentally from those of the document-centric internet infrastructure currently deployed at scale. While existing web protocols were designed to facilitate human-initiated document retrieval, the agentic paradigm necessitates infrastructure supporting autonomous entities that discover capabilities, negotiate terms, and coordinate actions at millisecond timescales across organizational boundaries. Contemporary agent deployments remain constrained within proprietary ecosystems - closed platforms where agents lack standardized discovery mechanisms, portable identity systems independent of platform ownership, and open protocols for cross-organizational transactions.

This architectural fragmentation mirrors the early internet's "walled garden" era, exemplified by services such as AOL that operated as closed networks with proprietary directories. The transition from closed platforms to open web infrastructure for documents provides a historical precedent for the infrastructure challenge facing the agentic web. However, the scale requirements differ substantially: the internet is projected to host not millions or billions, but eventually trillions of autonomous agents that negotiate, delegate tasks, and migrate between computational hosts in milliseconds - a load profile fundamentally distinct from human web usage patterns.

Project Nanda represents an open research initiative originating from MIT Media Lab that develops foundational infrastructure for an internet of AI agents. The project delivers four core technical components: the Nanda Index for discovery, registries for identity management, open coordination protocols, and Nanda Town for simulation-based validation. This work establishes technical mechanisms enabling agents from disparate organizations to discover each other's capabilities, establish trust relationships, delegate work, conduct transactions, and learn across organizational boundaries without requiring permissions from centralized authorities. This analysis examines the architectural principles, technical mechanisms, and validation methodologies underlying Project Nanda's approach to building open infrastructure for the agentic web.

2. Background and Related Work

2.1 Agent Architecture and Operational Requirements

An agent is formally defined as a model that employs tools within an iterative control loop. Given an objective, the agent determines subsequent actions, invokes appropriate tools, observes execution results, and continues this cycle until task completion. This fundamental loop serves as the architectural foundation upon which higher-order capabilities - including memory systems, orchestration frameworks, and multi-agent coordination mechanisms - are constructed. For agents performing consequential operations, access to production tools and applications becomes essential, creating requirements for robust control mechanisms, operational transparency, and clear ownership attribution.

The operational context for agents differs qualitatively from traditional web services. Self-hosted, open-source agent implementations prove critical because they provide users with control over three fundamental dimensions: who controls the agent's decision-making processes, where computational execution occurs, and the degree of visibility into operational behavior. When agents possess access to real tools capable of executing consequential actions, these control and transparency properties become essential rather than optional features.

2.2 Infrastructure Limitations and Scale Requirements

Existing internet infrastructure, particularly the Domain Name System (DNS), was architected primarily for name-to-address mapping of static resources. This design proves inadequate for agent-scale requirements where discovery must encompass not merely network location, but also operational capabilities, permission structures, behavioral rules, and communication protocol specifications. The DNS model - mapping human-readable names to IP addresses - addresses only a fraction of the metadata agents require to establish productive interactions.

The anticipated operational scale exacerbates these limitations. Current systems strain under agent-scale requirements because the load characteristics differ fundamentally from document retrieval patterns. Trillions of autonomous agents negotiating and delegating tasks in milliseconds generate traffic patterns, discovery queries, and coordination requirements that exceed the design parameters of document-centric infrastructure. Furthermore, contemporary agent ecosystems operate within proprietary platforms that provide discovery and coordination exclusively within platform boundaries, preventing cross-organizational collaboration and creating structural dependencies on platform-specific identity and trust mechanisms.

3. Core Analysis

3.1 Three-Layer Infrastructure Architecture

Project Nanda implements a three-layer architectural model comprising discovery, commerce, and bazaar layers. The discovery layer, instantiated through the Nanda Index, provides agents with a shared registry for publishing identity, capabilities, and network reachability information. Unlike DNS, which maps names to addresses, the discovery layer must answer qualitatively different queries: what does an agent do, what tools can it employ, what operational rules govern its behavior, and through what protocols can it be reached.

The commerce layer establishes portable identity and trust mechanisms that function independently of any single platform's ownership. This layer implements cryptographically-signed agent facts records that contain agent identity, capabilities, permissions, builder attribution, and reachability information. These records enable trust verification before agent connections are established, providing a foundation for cross-organizational coordination without requiring central authority validation.

The bazaar layer provides open protocols enabling agents to coordinate, transact, and learn across organizational boundaries. This layer addresses the fundamental challenge that "the hard problems of the agent web live between agents at scale" - specifically, how thousands of agents discover each other, prove identity, decide whom to trust, and coordinate without central authority. The bazaar metaphor emphasizes decentralized, peer-to-peer coordination rather than hierarchical control structures.

3.2 Agent Discovery and Identity Mechanisms

The Nanda Index implements an agent identity format using email-like structure: agent@domain.com. This standardization provides familiar syntax while enabling domain-based attribution and verification. When queried, the Index returns an agent card containing three critical elements: what the agent is (capabilities and purpose), how to reach it (network endpoints), and where to send messages (message routing information).

Message routing incorporates an intermediate message box layer that performs several critical functions before messages reach agents. This layer checks sender identity, implements access control policies, filters spam and malicious requests, and queues messages until the receiving agent is ready to process them. This architecture separates message receipt from message processing, enabling agents to manage computational resources efficiently while maintaining availability for discovery.

The adaptive resolution mechanism represents a significant architectural innovation. Rather than returning static endpoint information, the system returns updated agent facts based on request context, routing traffic to optimal endpoints and protecting private details based on requester identity and permissions. This context-aware resolution enables sophisticated traffic management, privacy protection, and endpoint selection without requiring centralized coordination.

3.3 Agent Onboarding and Hosting Infrastructure

Project Nanda provides three distinct onboarding pathways designed to accommodate entities ranging from large enterprises to individuals. Enterprises can operate their own agent catalogs and register gateways from owned domains, maintaining full infrastructure control. Existing websites can leverage DNS AID to connect agents to already-owned domains, reducing infrastructure requirements while maintaining domain-based attribution. Individuals and small businesses can utilize host39.org, a hosted service providing agent URLs without requiring domain ownership, lowering barriers to participation.

Agent hosting presents a fundamental trade-off between control and operational burden. Local hosting provides complete control but requires users to maintain uptime and availability. Cloud hosting proves more practical for most deployment scenarios, with options including general cloud platforms (AWS, Azure) for enterprise-ready solutions or agent-specific platforms (Maritime) optimized for agent workload characteristics.

The sleep/wake architecture implemented in platforms like Maritime addresses a critical economic constraint in multi-agent deployments. Idle agents do not consume computational resources, dramatically reducing per-agent operational costs. This architectural pattern becomes essential when running many agents simultaneously for teams, products, or large-scale simulations, transforming the economics of agent deployment from prohibitively expensive to practical.

3.4 Simulation Infrastructure and Protocol Validation

Nanda Town provides an open-source simulation environment for testing agentic infrastructure before real-world deployment. The system implements discrete event simulation methodology: scenarios are defined in YAML configuration files, agents and traffic patterns are injected into the simulation, and end-to-end results are measured systematically. This approach enables reproducible testing and protocol validation without requiring production-scale infrastructure.

The simulation architecture decomposes the agentic web into twelve distinct components: transport, communication, identity, registry, authentication, trust, payments, coordination, negotiation, memory, privacy, and data effects. This modular decomposition enables independent testing of individual layers before integration, facilitating systematic validation and iterative refinement. The system provides visual mapping of agent locations, real-time message movement visualization, protocol comparison capabilities, and step-by-step replay functionality for detailed analysis.

Nanda Town implements a two-tier testing approach. Tier one employs simple scripted agents for protocol validation and infrastructure testing, enabling rapid iteration without the computational overhead of full language models. Tier two swaps in actual AI models, testing how real agents behave under production-like conditions. This tiered approach balances testing thoroughness with computational efficiency, enabling comprehensive validation across multiple scales of complexity.

4. Technical Insights

The architectural patterns implemented in Project Nanda reveal several technical insights with broader applicability to distributed agent systems. The email-like agent identity format (agent@domain.com) provides a familiar, human-readable syntax while enabling domain-based verification and attribution. This design choice reduces cognitive overhead for developers and users while maintaining technical rigor for programmatic interactions.

The intermediate message box layer represents a critical architectural decision that separates message receipt from message processing. This separation enables sophisticated access control, spam filtering, and resource management without requiring agents to remain continuously active. The pattern proves particularly valuable in scenarios where agents must be discoverable and reachable but should not consume computational resources when idle.

Cryptographically-signed agent facts records establish a foundation for decentralized trust without requiring central certificate authorities or platform-specific verification mechanisms. This approach enables portable identity that follows agents across platforms and organizational boundaries, addressing a fundamental limitation of platform-specific identity systems. However, the effectiveness of this mechanism depends critically on key management practices and the security of signing infrastructure.

Adaptive resolution protocols introduce context-aware endpoint selection based on requester identity and permissions. This mechanism enables sophisticated traffic routing, privacy protection, and endpoint optimization without centralized coordination. The trade-off involves increased complexity in resolution logic and potential attack surfaces if access control policies are misconfigured. Implementation requires careful consideration of policy specification languages and verification mechanisms.

The sleep/wake architecture for agent hosting addresses computational efficiency in multi-agent deployments by eliminating resource consumption for idle agents. This pattern proves essential for economic viability when running hundreds or thousands of agents simultaneously. However, it introduces latency considerations - agents must wake before processing messages, creating potential delays in time-sensitive coordination scenarios. System designers must balance efficiency gains against responsiveness requirements based on application characteristics.

5. Discussion

The infrastructure patterns developed in Project Nanda establish foundational protocols for transitioning from proprietary agent platforms to an open agentic web. The architectural parallels to the early web's evolution from walled gardens to open infrastructure prove instructive: just as the open web required shared protocols for resource identification (URLs), content transfer (HTTP), and hypertext formatting (HTML), the agentic web requires shared protocols for agent discovery, identity verification, and coordination. Project Nanda's three-layer architecture provides these foundational protocols while maintaining architectural openness for innovation at higher layers.

Several areas warrant further investigation. The security model for cryptographically-signed agent facts records requires robust key management infrastructure and revocation mechanisms. The effectiveness of spam filtering and access control in the message box layer remains to be validated at internet scale, where adversarial actors will attempt to exploit discovery and messaging infrastructure. The economic models for agent hosting and coordination require development - while sleep/wake architectures improve computational efficiency, questions remain regarding pricing models, resource allocation, and incentive alignment in decentralized agent networks.

The simulation infrastructure provided by Nanda Town addresses a critical gap in agent system development: the ability to test coordination protocols before deployment at scale. However, the fidelity of discrete event simulation for modeling real agent behavior merits careful consideration. The two-tier approach - scripted agents for protocol testing, real models for behavior testing - provides a pragmatic balance, but questions remain regarding how well simulation results predict production behavior, particularly in adversarial scenarios or under resource constraints not modeled in simulation.

The broader implications extend beyond technical architecture to governance and coordination models for shared infrastructure. Open infrastructure requires maintenance, security updates, and evolutionary development without centralized control. The mechanisms for coordinating infrastructure evolution, resolving disputes, and managing shared resources in a decentralized agent web remain open research questions. Historical precedents from internet governance provide partial guidance, but the autonomous nature of agents introduces novel challenges in accountability, liability, and collective decision-making.

6. Conclusion

Project Nanda establishes foundational infrastructure for an open, decentralized agentic web through a three-layer architecture comprising discovery mechanisms, portable identity protocols, and open coordination frameworks. The Nanda Index provides shared discovery infrastructure enabling agents to publish and query capabilities across organizational boundaries. Cryptographically-signed agent facts records establish portable identity independent of platform ownership. The message box architecture separates message receipt from processing, enabling sophisticated access control and resource management. Nanda Town provides simulation infrastructure for protocol validation before production deployment.

The practical implications prove substantial for organizations developing agent systems. Rather than building within proprietary platforms that limit cross-organizational coordination, developers can leverage open protocols for discovery, identity, and messaging. The sleep/wake architecture demonstrates that multi-agent deployments can achieve economic viability through architectural patterns that eliminate idle resource consumption. The modular decomposition of agentic infrastructure into twelve distinct components provides a framework for systematic development and testing.

Future work should address key management infrastructure for signed agent facts, validation of spam filtering and access control at internet scale, economic models for decentralized agent coordination, and governance mechanisms for shared infrastructure evolution. As the number of deployed agents grows from millions toward the projected trillions, the foundational protocols established by Project Nanda will prove critical for enabling the transition from fragmented, proprietary agent platforms to an open agentic web capable of supporting cross-organizational coordination at machine timescales.

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub