September 25, 2025
In Part 1 of this series, we introduced Agent Oriented Architecture (AOA) as a conceptual framework for the AI-native future. We contrasted today’s “faster horses”—AI bolted onto human workflows—with the “automotive age”: AI-native systems where intelligent agents are first-class citizens. The promise of AOA lies in emergent collaboration, where autonomous agents dynamically discover, negotiate, and cooperate to achieve complex objectives without a rigid flowchart.
But this promise hinges on a single critical question: How do agents find each other? The mechanism of discovery defines the capabilities, limitations, and ultimately the intelligence of the system. Without a robust answer, AOA remains a theory rather than a working architecture.
This article bridges theory and practice. We explore a pioneering real-world initiative from MIT—Project NANDA—and then examine how AOA adapts these ideas to enterprise needs through the design of the Agent Registry.
It’s validating to see leading institutions explore parallel ideas. MIT’s Project NANDA is a landmark effort to build the foundational infrastructure for an “Internet of AI Agents.” NANDA is not a competitor to AOA, but a powerful validation of the paradigm: the challenges of discovery, communication, and trust are exactly what researchers are tackling to support trillions of agents worldwide.
NANDA’s core components provide a strong reference model:
NANDA’s focus is clear: building the public square, the internet-scale infrastructure of identity, routing, and trust. Enterprises can build on this, but their needs go deeper—toward intelligence and value creation inside the organisation.
AOA and NANDA share a conceptual foundation. Both rely on modular, autonomous agents that discover one another dynamically, and both embrace standards like A2A for communication and MCP for tool access. This convergence might signal a broader shift in the software industry.
But, their destinies diverge:
This divergence shows up most clearly in their registries:
This is not a subtle distinction—it requires a fundamentally different technological solution.
The Agent Registry is the cornerstone of AOA. Unlike NANDA’s metadata-driven directory, the AOA Registry must handle intent-based queries—open-ended requests where the agent doesn’t know exactly which capabilities are needed.
Traditional document databases (e.g. MongoDB) can store structured metadata and support filtering or hybrid keyword searches: “Find all finance agents owned by the EMEA department.” Useful, but insufficient when queries demand semantic understanding, like “Correlate logistics delays with customer satisfaction scores.”
Here, metadata breaks down. The registry needs to understand meaning, not just match fields. That is why a vector database is non-negotiable for AOA.
Vector embeddings: AI-generated numerical representations capture the semantic meaning of text or images. Phrases like “quarterly earnings report” and “summary of financial performance this period” map to nearly identical vectors despite different words.
Similarity search: Vector databases use algorithms like Approximate Nearest Neighbour (ANN) to find semantically “close” results, enabling discovery by meaning rather than keywords.
Applied to AOA, this transforms the registry:
Crucially, this creates evolutionary pressure against the super-agent trap. A monolithic “do-everything” agent produces a diluted vector, while specialised agents produce precise embeddings. When queries arrive, specialized vectors win, naturally rewarding diversity and scalability.
| Feature | NANDA-style Index (Metadata-centric) | AOA Registry (Semantic-centric) |
| Primary technology | Document DB with vector add-on (e.g., MongoDB Atlas) | Purpose-built vector DB (e.g., Milvus, Pinecone, Weaviate) |
| Query method | Keyword search, metadata filtering, hybrid search | Semantic similarity search (ANN) |
| Discovery paradigm | Lookup: “Find the agent named Q3_Report_Gen.” | Discovery: “Find an agent that can synthesize financial + logistics data.” |
| Core strength | Identity, routing, verifiable credentials, structured filtering | Intent matching, capability discovery, handling ambiguity |
| Analogy | Secure corporate directory (LDAP) | Council of experts you can query with a complex problem |
NANDA builds the global public square. But within the enterprise, AOA demands more: a semantic Agent Registry powered by vector databases. This is not a technical nuance—it’s the difference between systems that follow commands and platforms that understand intent.
By anchoring AOA in semantic discovery, enterprises may gain replaceability of components, scalability of capabilities, and a resilient ecosystem of specialised agents. This registry is the cornerstone of emergent, adaptive, intelligent systems.
With discovery in place, the next challenge emerges: scaling collaboration. In Part 3, we’ll explore how agents talk to each other, negotiate tasks, and coordinate multi-step workflows. Just as important, we’ll examine what makes these systems scale: the role of Small Language Models (SLMs) in powering efficient, specialised agents.
SLMs are efficient, easier to fine-tune, and perfectly aligned with AOA’s principle of specialisation. Small is going to be huge—and we’ll show how SLMs unlock resilient, scalable agentic workflows while keeping human oversight at the centre.
All business models have their pros and cons. But, when you consider the type of problems we help our clients to solve at Equal Experts, it’s worth thinking about the level of experience and the best consultancy approach to solve them.
If you’d like to find out more about working with us – get in touch. We’d love to hear from you.
