AgNet: A Novel AI Agent Network Architecture

Manoj Gupta Vikram Acharya Sai Sujan Ayush Mittal

[email protected] [email protected] [email protected] [email protected]
Plotch.ai Plotch.ai Plotch.ai Plotch.ai

Abstract—A big agentic switch is happening globally due to AI to enhance organizational productivity [6].
(Artificial Intelligence) whereby consumers and enterprises are
switching to AI agents for task completions in an autonomous
The influence of Agentic AI on productivity within
way. This paradigm shift due to AI is leading to proliferation
of AI agents in enterprises requiring innovative agent network enterprises is significant. By automating repetitive and time-
architectures for better discovery, communication and security of consuming tasks, organizations can shift human resources
these agents. Here we present a novel agent network architecture to focus on strategic and creative initiatives. Additionally,
where we introduce innovative concepts like agent registry, agent Agentic AI supports better decision-making via data-driven
name server, agent text transfer protocol (ATTP), agent private
insights, boosts operational efficiency, and promotes scalability
cloud and agent gateway. Quite a lot of parallels have been
drawn from current world wide web (www) Internet architecture in workflows. With organizations progressively implementing
where similar components exist to support discovery, security and networks that incorporate millions of AI agents, the necessity
communication. for resilient architectures such as AgNet becomes essential to
Index Terms—AgNet, AI Agent Network Architecture, AI guarantee smooth communication, security, and scalability.
Agents, Agent Discovery, Agent Registry, Agent Name Server
(ANS), Agent Text Transfer Protocol (ATTP), Agent Private
Cloud (APC), Zero Trust Architecture (ZTA), Agent Security, This change in approach requires creative architectures that
Communication Protocol, Scalability, Interoperability. can absorb many AI agents into a single ecosystem [9]. The
major challenges are secure communication, dynamic agent
I. I NTRODUCTION discovery and operational scalability [8]. AgNet introduces
A crucial component of contemporary technology, artificial a new framework to address these needs, inspired by the
intelligence (AI) is spurring innovation and revolutionizing principles that govern internet protocols and infrastructure.
sectors all over the world. Fundamentally, AI is the process
by which machines mimic human intellect in order to carry The AgNet framework presents essential elements like the
out functions like learning, thinking, and decision-making. Agent Registry, Agent Name Server (ANS), Agent Text Trans-
AI agents are among the most revolutionary developments fer Protocol (ATTP), and Agent Private Cloud (APC). Every
in AI; they are self-governing software systems created to one of these elements is crafted to replicate features similar
interact with their surroundings, make choices, and carry out to the World Wide Web, modified for the specific needs of
activities with little assistance from humans. AI agent networks. This blend of scalability, composability,
and security renders AgNet a strong solution for the changing
The proliferation of AI agents within enterprises has ush- environment of agentic AI.
ered in a new era of productivity and automation, redefining
traditional workforce dynamics. The number of AI agents in III. N ETWORK A RCHITECTURE
large organizations may exceed millions which will create
challenges in task performance, deployment, security and AgNet is built on four principles: efficiency, security,
scalability. To manage such large network of AI agents in any scalability and composability. By integrating key subsystems
organization would require a novel network architecture which for robust agent detection, communication and security you
would allow such a network to function seamlessly. Here we can realise these.
present AgNet which is a novel network architecture for AI
agents. Modular design means you can add new agents and
subsystems without affecting the core of the network.
II. BACKGROUND By breaking the architecture into separate but connected
AI agents are essential for optimizing processes and components AgNet allows for adaptability and agility to
boosting efficiency in multiple areas, ranging from customer support small and large networks of AI agents. Plus modular
service bots to advanced data analytics systems. These design means you can deploy to specific organisations or
agents can operate both independently and as members sectors.
of multi-agent systems, where their cooperation facilitates
the completion of complex, interrelated tasks [1]. This Security is built in to ensure safe operation at every level.
development has led to Agentic AI, a concept that sees AI Zero trust for authentication and authorisation and strong
agents as collaborators, smoothly blending with human teams encryption for data in transit and at rest. AgNet is interoperable
so agents can talk to each other across different systems and box, addressing this critical issue of interoperability between
older protocols. agents. Its a very important part of agent communication.
c) Role in Agent Discovery: Communication which is
achieved using the agent UUID plays key role in communi-
cation and agent discovery. So in a system with many agents,
wrong one shouldn’t get called, above all, the call is quickest
and secure. The agent discovery is performed using the Agent
UUID wherein every agent is given a unique UUID that is
mapping over cryptographic tokens and metadata resulting in
better agent authentication and secure transactions among the
entities of the network.
B. Agent Registry
The Agent Registry is the central database where all the
metadata for every agent in the AgNet ecosystem is stored.
Fig. 1. Agent Network Architecture This is an important layer of the ecosystem. Plays a key role
in agent identification, capability management and lifecycle
monitoring.
A. Agent Discovery
a) Architecture of the Agent Registry: The architecture
Agent discovery is a key part of the AgNet framework to of the agent registry is designed to make it easy to handle
enable safe and efficient inter-agent interactions in a large and access agent metadata in the AI Agent Network (AgNet).
network of AI agents. When agents can be in the millions, At the base of it is a Metadata Repository which is a
strong agent identification, resolution and communication is central database to store all the metadata of every agent.
crucial. This section will cover the key components of the It stores key information such as the UUID, the agent’s
agent discovery process—Agent UUID, Agent Registry and capabilities and functions, its current status and availability
Agent Name Server (ANS)—and how they help in seamless and the communication endpoints (urls and APIs) of the agent.
communication in the AgNet ecosystem. This central repository ensures consistency and reliability of
metadata.
To keep the system up to date the registry has Dynamic
Updates so changes can be made in real time. Whether
agents are launched, updated or retired the registry reflects
these changes instantly so the network is always current and
working.
b) Key Functionalities of the Registry: The agent registry
fulfills several pivotal functions to maintain the integrity and
efficiency of AgNet:
• Agent Discovery: The registry is the backend for the
Agent Name Server (ANS). It resolves agent queries and
connects users to the right agents using metadata.
Fig. 2. Agent Discovery Framework • Capability Matching: By analyzing the capabilities reg-
istered for each agent, the registry matches incoming
a) Agent UUID: The Agent Universally Unique Identi- queries or tasks to the most suitable agents, ensuring
fier (Agent UUID) is very important in AgNet framework, a optimal resource allocation and task execution.
unique and immutable identifier for every AI agent globally. • Agent Lifecycle Management: The registry follows the
This identifier helps in identifying, interacting and monitoring entire lifecycle of agents from creation and deployment
agents across the network, a solid foundation for agent based to modification and decommissioning. This ensures meta-
systems. data is consistent across the network and reduces errors
b) Design Principles: The Agent UUID is optimized and inefficiencies.
to fulfill three main criteria to be performant and reliable. • Integration with the Agent Name Server: The registry
Top three criteria are: worldwide uniqueness, durability and is a data source for the ANS. When a query is received the
standard compliance, — all of them are essential. Uniqueness ANS looks up the registry to get the UUID and endpoint
guarantees that no UUID will ever be repeated in the network. of the agent and then communicates and executes the
Durability makes sure that the UUID continues to remain the task.
same across the agent’s life time, which helps in long term c) Advantages of the Agent Registry: The agent registry
tracking, auditing and accountability. You have a standard- is good for the whole AgNet architecture. It’s a central
compliant system, which means that it will work out of the database from call to managing agents. It’s the base of the
whole AgNet system. Very important for infrastructure as this protocol gaps and adds features.
is one of the first point of agent discovery and workflow.
To sum up, the agent registry is a part of the AgNet frame- This section will go into detail about the communication
work, metadata handling, agent identification and lifecycle protocols and components that make agent to agent commu-
management. Its connection to the Agent Name Server and nication reliable, scalable and secure.
real time updates means the network is scalable, consistent
and adaptive to the changing world of AI.
C. Agent Name Server (ANS)
The Agent Name Server (ANS) is the discovery engine in
the AgNet architecture and acts as an intermediary that maps
user intents to agents. It hides the complexity of the discovery
process so Client-Agents can find and talk to Server-Agents.
a) Core Responsibilities: ANS is important for intent
resolution, caching and fallback. Intent resolution is the pro-
cess of taking high level intents from Client-Agents and
turning them into actionable endpoints by querying the Agent
Registry. Caching allows ANS to keep a local cache of
Fig. 3. Agent Communication Framework
frequently used agent information so performance and delay is
minimized. When cache misses happen the fallback system al- 1) Agent Text Transfer Protocol (ATTP): The Agent Text
lows ANS to actively query the Agent Registry and get/refresh Transfer Protocol (ATTP) is AgNet fabric’s agent to agent
the required data. communication supervisor. It aims to be a light, secure, and
b) Query Flow: The query process begins when a Client- efficient means of transporting data at low latency and high
Agent sends a query with the intent and entities to the ANS. standards of safety.
ANS checks its cache for the Agent UUID and endpoint. If the a) Key Features: It is high performance, low latency, and
cache doesn’t have the data, ANS queries the Agent Registry can scale to millions of simultaneous interactions. Security
and gets the data. Then the UUID and endpoint is sent back is a fundamental quality, and encryption, authentication and
to the Client-Agent. tokenization are features baked into the protocol to ensure
c) Integration with ATTP: The ANS is tightly coupled the secure transfer of data. This modular process makes the
with the ATTP protocol to enable secure communication protocol easy to implement as you can plug in new features
between agents. All queries and responses are sent as ATTP or extend agent capabilities.
messages that include encryption, authentication and version- b) Protocol Structure: ATTP message consists of two
ing headers. This integration provides a seamless and secure components: header and body. The header includes version,
interaction framework within the AgNet. message type, agent-specific identifiers such as Agent Process
d) Security Features: To have a safe operational environ- Control ID (APCID) and Agent UUID, authentication tokens,
ment the ANS has various security features. Authentication timestamps, and encryption flags as metadata. Data like the
is done by verifying the identity of the querying Client- payload carries the body or the query or task to be processed
Agents through authentication tokens in the ATTP headers. along with relevant metadata such as identified intent and
Access control ensures only authorized users can search or entities. A diagram of this is shown in Figure 4.
retrieve sensitive data. Data integrity is maintained through
cryptographic methods to prevent data from being modified
during transmission.
e) Advantages: ANS simplifies the discovery process
for Client-Agents by abstracting away direct interaction with
the registry. Its caching and resolution mechanisms improve
overall performance of the AgNet. Fallback mechanisms and
robust error handling capabilities improve fault tolerance so
the system can work under various conditions.
D. Agent Communication
Agent communication in AgNet is an advanced, secure
and efficient way for AI agents to talk to each other across
the network. This is done through the Agent Text Transfer Fig. 4. Agent Communication Framework
Protocol (ATTP) which is the heart of the system. To ensure
compatibility with different environments and older systems c) Communication Flow: The communication flow in
the architecture has the ATTP Adaptor which bridges ATTP involves several steps to process the request correctly.
First the Client-Agent extracts the intent and entities from the
user query. The extracted message is then sent to the Server-
Agent through the Agent Name Server (ANS). The Server-
Agent processes the request and sends the response back to
the Client-Agent and that’s the communication cycle.
d) Advantages: ATTP has many benefits for the Ag-
Net ecosystem. It facilitates more seamless interaction, as it
provides standardized communication to better integrate as
a whole system between disparate agents. Intelligent routing
and caching are also leveraged to optimize details and reduce
latency in data exchange. And notion has strong security
measures like encryption and authentication that maintain the Fig. 5. Agent Security Framework
integrity and confidentiality of data.
2) ATTP Adaptor: The ATTP Adaptor bridges the agents
and external systems by translating incoming payloads, ex- requests to ensure security protocols are met, encrypts
tracting intent and entities and routing through the ANS. communication and routes authenticated requests to the
a) Role in Workflow: The ATTP Adaptor plays a key internal agents. This ensures only authorized and legiti-
role in processing incoming requests by extracting the routing mate requests are processed.
information. It uses the Agent Name Server (ANS) to resolve • Data and Task Isolation: To reduce risk the APC uses
the destination agent and route the query. And it integrates with resource segmentation, which puts agents and their tasks
the orchestrator to make sure the data fits the overall workflow in separate areas. These areas are defined by roles and
of the AgNet system. permissions so unauthorized access is prevented and the
b) Advantages: ATTP Adaptor Benefits. It provides in- scope of any breach is limited.
teroperability by enabling communication between different • Real-Time Surveillance and Anomaly Identification:
agents and protocols. Due to its scalable nature, it is easy The APC watches agent actions and task performance to
to add new agents or tools to the network. It also comprises detect any abnormalities or signs of security breaches.
robust error mitigation, which gracefully handles both erro- Machine learning algorithms detect anomalies like fre-
neous inputs and routing errors, ensuring the system operates quent login failures or unusual task behavior so threats
smoothly within the AgNet environment. can be managed proactively.
The ATTP Adaptor is an important component that adds • End-to-End Encryption: All data in the APC is en-
flexibility and scalability to the AgNet ecosystem. crypted whether it’s being sent or stored. This ensures
confidentiality and integrity of sensitive information like
E. Agent Security
user details, task specs and operational details, blocks
Agent security [11] [12] is a fundamental part of the AgNet unauthorized access and maintains data integrity across
architecture to ensure the safe and reliable operation of its the system.
distributed AI agent ecosystem. In a highly dynamic and
The APC follows the principles of Zero Trust Architecture
interactive environment, keeping the confidentiality, integrity
(ZTA). Continuous authentication, minimal privilege access
and availability of communication and data is critical [7]. To
and flexible security is applied throughout the environment.
address this AgNet uses Zero Trust Architecture (ZTA) [10]
By limiting agents to only the resources they need for their
principles which is ”never trust, always verify”. The Agent
job the APC reduces the attack surface so even if an agent is
Private Cloud (APC) is the mechanism to implement these
compromised it’s not a big threat to the system.
principles, it provides a secure and isolated environment for
all agents.
1) Agent Private Cloud (APC): The Agent Private Cloud Furthermore, AgNet’s overall security system which in-
(APC) is the base of AgNet’s security framework. It’s a cludes ANS, Agent Registry and ATTP has the APC. These
secure sandbox that enforces strict security protocols and components together form a layered security system. By ensur-
operational isolation so all agent interactions happen in a ing that every action in the APC is verified and controlled, this
controlled and monitored environment. The APC architecture integration supports AgNet’s goal of being robust, scalable,
has advanced security features that prevent unauthorized and trusted across the distributed agent network.
access, data breaches and malicious activity so operations can
run smoothly and securely. IV. R ESULTS

The APC incorporates several key functionalities to achieve We have validated AgNet in both simulations and real-
these security goals: world applications to demonstrate its relevance, safety, and
• Gateway Agent: The Gateway Agent is the entry point operational efficiency. Key findings from the assessment are
for external requests into the APC. It validates incoming detailed below:
A. Applicability Assessment R EFERENCES
The AgNet system was successfully evaluated using a [1] T. Masterman, S. Besen, M. Sawtell, and A. Chao, The Landscape of
network of four online commerce agents. Every agent has Emerging AI Agent Architectures for Reasoning, Planning, and Tool
Calling: A Survey, arXiv preprint arXiv:2404.11584v1, 2024.
registered and found each other successfully communicates [2] J. S. Rosenschein and G. Zlotkin, Infrastructure for Discovery in Multi-
using natural language via the ATTP format and executes agent Systems, Technical Report CMU-CS-99-123, Carnegie Mellon
tasks like perform product searches, respond to customer order University, 1999.
[3] O. Shehory and S. Kraus, Methods for Task Allocation via Agent
status inquiries, and generate catalogs for sellers. In addition, Coalition Formation, Artificial Intelligence, vol. 101, no. 1-2, pp. 165-
ANS reduced the average agent lookup latency by 65% due 200, 1998.
to its caching and fallback systems. [4] F. Bellifemine, A. Poggi, and G. Rimassa, Developing Multi-agent Sys-
tems with a FIPA-compliant Agent Framework, in Software Engineering
B. Security Performance for Large-Scale Multi-Agent Systems, P. Ciancarini and M. Wooldridge,
Eds., Springer, 2001, pp. 203-217.
The Agent Private Cloud [APC] proved to sufficiently [5] E. Johnson, Agent Communication Protocols, Smythos Artificial Intel-
contain bad agents and did not allow access to key network ligence, Mar. 2024.
[6] T. Abuelsaad, D. Akkil, P. Dey, A. Jagmohan, A. Vempaty, and
components. In addition, the end-to-end encryption added in R. Kokku, Agent-E: From Autonomous Web Navigation to Foundational
the ATTP protocol did not result in data breach during simu- Design Principles in Agentic Systems, arXiv preprint arXiv:2407.13032,
lated Man-In-The-Middle (MITM) attack scenarios, showing 2024.
[7] R. S. Hallyburton, D. Hunt, S. Luo, and M. Pajic, A Multi-Agent Security
that security was well and truly up to scratch. Testbed for the Analysis of Attacks and Defenses in Collaborative Sensor
Fusion, arXiv preprint arXiv:2401.09387, 2024.
C. Operational Efficiency [8] G. Brown, Introduction to LMOS, Eclipse Foundation, May 2024.
Agent Registry facilitated task allocation between these [9] H. Lee and I. Kim, Multi-agent Systems: A Survey about its Components,
Framework, and Workflow, International Journal of Advanced Computer
lines with a success rate of 92% by mapping agent capabilities Science and Applications, vol. 15, no. 3, pp. 45-58, 2024.
to task requirements, astronomically improving task efficiency. [10] D. Smith, Understanding AI Risks and How to Secure Using Zero Trust,
ATTP’s modular approach minimized the protocol overhead LevelBlue Security Essentials Blog, Apr. 2024.
[11] P. Novák, M. Rollo, J. Hodı́k, and T. Vlček, Communication Security
by 25%, leading to faster transmission than traditional agent in Multi-agent Systems, in Multi-Agent Systems and Applications III,
communication protocols. However, the AgNet architecture V. Mařı́k, J. Müller, and M. Pechouček, Eds., Springer, 2003, pp. 454-
can effectively optimize communication and task processing, 463.
[12] A. Baudet, O. E. Aktouf, A. Mercier, and P. Elbaz-Vincent, Systematic
as shown by the efficiency gain. Mapping Study of Security in Multi-Embedded-Agent Systems, IEEE
Access, vol. 9, pp. 160-175, 2021.
V. C ONCLUSION
This paper introduced AgNet, an AI agent network
architecture for large-scale AI agent environments.
AgNet addresses the problems of scalability, security
and communication by introducing the Agent Registry, Agent
Name Server (ANS), Agent Text Transfer Protocol (ATTP),
and Agent Private Cloud (APC).

These findings show that AgNet can control networks

consisting of a few agents with good performance in terms
of security and system performance. The architectural
scalability makes it suitable for organizations of all sizes,
and the APC’s embedded security measures provide trust
and resilience against the ever-evolving cyber threat landscape.

In the future, we will focus on testing the network with

millions of agents, extend the ATTP protocol (for cross-
network interoperability) and add advanced AI models to
enable dynamic capability matching and anomaly detection.
AgNet represents a significant step towards grouping the
potential of agentic AI to optimize efficiency and automate
organizations.
ACKNOWLEDGMENT
We acknowledge the amazing effort of the extended research
team at Plotch.ai who tirelessly worked on various aspects
of the Agentic research. These extended researchers included
Monica Gupta, Neha Vats, Tarun Nagpal, Tarun Choudhary,
Shrey Pant, Vijay Balaram and Karthikeyan M.