Research Overview
I'm passionate about designing the foundational infrastructure for the Internet of AI Agents — a paradigm where autonomous software agents can interact, negotiate, and collaborate across digital systems without relying on continuous human input. My work explores how to make this transition possible by building decentralized, verifiable, and privacy-preserving protocols for agent discovery, communication, and trust.
Today's internet was designed for static web pages and human-triggered APIs. As we move toward a world with trillions of intelligent agents — coordinating logistics, automating workflows, conducting research — the underlying assumptions of the web must be rethought. We need new abstractions that support dynamic identity, real-time coordination, and programmable trust.
Through Project NANDA, I've coauthored two core papers that examine both the need for — and the architecture of — this new agentic infrastructure.
Beyond DNS: Unlocking the Internet of AI Agents via the NANDA Index and Verified AgentFacts
The Internet is poised to host billions to trillions of autonomous AI agents that negotiate, delegate, and migrate in milliseconds, creating workloads that will strain DNS-centered identity and discovery. This paper describes the NANDA index architecture as a means for discoverability, identifiability and authentication in the internet of AI agents.
We present an architecture where a minimal lean index resolves to dynamic, cryptographically verifiable AgentFacts that supports:
- Multi-endpoint routing, load balancing, and privacy-preserving access
- Credentialed capability assertions with schema validation
- Sub-second revocation and key rotation for trillions of entities
Our architecture design delivers five concrete guarantees:
- A quilt-like index proposal supporting both NANDA-native and third-party agents
- Rapid global resolution for newly spawned AI agents
- Privacy-preserving discovery across organizational boundaries via verifiable, least-disclosure queries
The result is a lightweight, horizontally scalable foundation that unlocks secure, trust-aware collaboration for the next generation of the Internet of AI agents, without abandoning existing web infrastructure.
Upgrade or Switch: Do We Need a Next-Gen Trusted Architecture for the Internet of AI Agents?
The emerging Internet of AI Agents challenges existing web infrastructure designed for human-scale, reactive interactions. Unlike traditional web resources, autonomous AI agents initiate actions, maintain persistent state, spawn sub-agents, and negotiate directly with peers—demanding millisecond-level discovery, instant credential revocation, and cryptographic behavioral proofs that exceed current DNS/PKI capabilities.
We identify critical failure points in current infrastructure:
- DNS propagation (24-48 hours vs. required milliseconds)
- Certificate revocation unable to scale to trillions of entities
- IPv4/IPv6 addressing inadequate for agent-scale routing
We evaluate three approaches for the Internet of AI Agents:
- 1Upgrade paths for existing infrastructure
- 2Switch options to purpose-built architectures
- 3Hybrid index/registries combining both approaches
Drawing parallels to dialup-to-broadband transitions, we find that agent requirements constitute qualitative, not incremental, changes. While upgrades offer compatibility and faster deployment, clean-slate solutions provide better performance but require longer adoption cycles.
Current Research Project
Emergent Behavior Analysis in Multi-Model Physical World Interactions
I'm investigating a fundamental question in AI systems: how do different large language models behave when granted agency over physical environments? This research examines the emergent behavioral patterns that arise when various AI models—including Grok, GPT-4, Gemini, and Claude—are given identical environmental contexts and allowed to autonomously determine their actions and interactions.
Experimental Framework:
Using OpenRouter as a unified interface, I present different models with standardized environmental descriptions: "You observe five individuals. Two are engaged in conversation, one is reading. A coffee cup, bicycle, and another robotic system are visible nearby." Each model generates structured JSON outputs specifying motor commands and verbal responses.
Preliminary Findings:
- Distinct behavioral signatures emerge across different model architectures
- Some models prioritize social interaction, others focus on object manipulation
- Behavioral patterns correlate with training data and architectural choices
- Physical embodiment significantly amplifies behavioral differences
Research Significance:
This work addresses critical gaps in understanding how AI systems will interact with physical environments as they become more autonomous. The findings have implications for robotics, human-AI collaboration, and the design of embodied AI systems. By examining behavioral emergence across different model architectures, we can better understand the relationship between training methodologies and real-world AI behavior.
This research represents a foundational step toward understanding how AI systems will behave when granted agency over physical environments—a critical question as we move toward more autonomous and embodied AI systems.
Agentic Sandbox Simulation
Large-Scale Multi-Agent Coordination Experiments
I'm building an agentic sandbox simulation where hundreds of autonomous agents running on different EC2 instances around the world communicate via agent-to-agent protocols. We're running large-scale experiments to understand how autonomous systems coordinate and collaborate at scale.
Core Research Questions:
- How can we solve coordination problems to make agents actually useful?
- Moving beyond random message exchanges that lead nowhere
- Creating meaningful agent interactions and task completion
Experimental Focus Areas:
- Agentic Web Indexing: How agents discover and catalog each other across distributed systems
- Intelligent Caching: Optimizing information sharing and storage across agent networks
- Transaction Management: Verifying transactions and ensuring reliable agent interactions (inspiration from web3 mempool)
- Distributed Coordination: Protocols for agents to work together on complex tasks
Technical Infrastructure:
Running hundreds of autonomous agents across multiple EC2 instances globally, connected via agent-to-agent communication protocols. Each agent operates independently while participating in coordinated experiments to test various coordination strategies and communication patterns.
This research builds directly on the NANDA index architecture and explores how autonomous agents will actually coordinate in the real world when deployed at scale.
Ongoing Research Interests
Random thoughts that keep me up at night...
→Agent-Human Collaboration: How will we actually communicate with them? How can we trust them? Is it JARVIS-like? Is it more like an iPhone or OpenAI listening device or a hologram? What does seamless collaboration between humans and autonomous agents actually look like in practice?
→Interface Layers: Will it be holograms or OpenAI listening devices or Neurolink? What's the actual interface through which we'll interact with AI agents? Are we looking at spatial computing, brain-computer interfaces, or something entirely different?
→Agentic Advertising: When we give our agents purchasing power, how can brands actually promote themselves? How do we ensure the market stays fair for smaller players? What happens when AI agents become the primary consumers instead of humans?
→Human Potential & Purpose: How do we help people achieve their human potential in a future where many will lose purpose and jobs, and games, VR, and agents become more prominent? What does meaningful work actually look like?
→Human-to-Human Connection: If AI companions become more reliable, empathetic, and available than human friends, what incentives remain for forming human relationships? This isn't just about loneliness—what happens to population dynamics and social structures?
→Education & Global Equality: The education system is clearly broken. How do we build systems to bring out potential in kids and educate them so they actually love learning? How can we make the world more equal regardless of what country or family you were born into? AI can definitely help, but there needs to be some coordinated global effort toward an equitable future.