
Almost a year ago, executives, researchers, and developers within the Outshift group of Cisco Systems – an incubation unit focused on such advanced technologies as AI and quantum computing – began batting about the idea of a network infrastructure connecting vast numbers of AI agents from multiple vendors or organizations, allowing those AI agents to automatically communicate, work together, and solve complex problems for enterprises.
The idea is that, as the nature of AI agents evolves and their use grows, there needs to be an open environment in which they can find each other and work together, regardless of what vendors built them. If an organization has a drug discovery project or sales forecasting problem that needs to be addressed, the AI agents they assign to it should be able to pull in other agents from other companies to solve it.
That’s not happening. Not yet, anyway. But it’s coming, according to Vijoy Pandey, senior vice president of the Outshift incubator, and it’s going to arrive in the form of the Internet of Agents.
“Largely, computing has been deterministic in nature for all these decades with probabilistic (stochastic) computing being relegated to special use cases (ML models) or quantum computing use cases,” Pandey wrote this month in a white paper on the subject. “Generalized probabilistic computing only took shape after the democratization of language foundation models such as ChatGPT in November 2022.”
The rise of foundation model-based AI-native applications, particularly in the form of AI agents, is happening quickly and will change how AI is used by enterprises and how work and business is done. Pandey pointed to the Internet and the Internet for clouds as examples of infrastructure adapting to rapid changes in business and IT environments. A similar shift is needed for AI agents.
“We stand at yet another inflection point – the move to distributed agentic computing,” he wrote. “One that will drive a platform paradigm shift larger than all other platforms in the past. This distributed agentic computation platform needs a new secure connectivity layer that sits above the current cloud native connectivity layer.”
The Internet of Agents will need to be fast, safe, and open, he wrote. It will involve “push-button hardware-software stacks” that include GPU and CPU virtualization, sophisticated pipelines for transmitting data, and tightly coupled compute clusters that provide high memory bandwidth and very low latency connections.
“This infrastructure must process agent collaboration at machine speed, potentially thousands of times faster than current,” Pandey wrote, adding that it also must be protected from upcoming quantum computing, which threatens encryption and other security methods. “This isn’t just about protecting data in transit – it’s about securing the entire chain of agent interactions across organizational boundaries.”
It’s also going to take industry-wide cooperation for such tasks as creating open standards for communication between agents, including defining how agents identify themselves, create trust, and securely share information. In additions, all companies should be able to implement the open protocols.
Organizations also will have to work together to create reference implementations for developers for security, scalability, and interoperability, a common infrastructure with open source tools for deploying, monitoring, and managing agents, and a security framework to protect against current threats and those coming with quantum computing.
There also should be a common way for enterprises to migrate to the new infrastructure.
AI companies could then offer their own technologies to the mix, including multi-tenancy, observability, and other enterprise features. Pandey said the data plane for Cisco’s implementation of the Internet of Agents would use its networking gear and extensions to open source tools for easier onboarding of agentic applications onto the infrastructure.
Outshift executives began publicly talking about the idea of an agentic AI-focused infrastructure last year, and in November 2024 the company published a research paper titled DAWN: Designing Distributed Agents in a Worldwide Network. But Cisco isn’t the only one exploring the Internet of Agents. Researchers with Tencent and several universities in China earlier published a research paper about the Internet of Agents, as did researchers from such institutions as Stanford University, Princeton University, and the University of California, Berkeley.
Pandey also noted talks by IBM and Amazon on the subject at a conference last year, and that startups like LangGraph and Lyzr also are moving in the space, thought he added that they’re “currently very far from it.”
Agentic AI has become a top focus of tech vendors and businesses over the past year. It’s the next step in generative AI, which puts a focus on LLMs that are trained on vast amounts of data being able to respond quickly in written content, images, and videos to prompts. AI agents are different in that they operate autonomously without much human intervention to solve complex problems. They can reason, plan, interact with their environments, collect data, and work with other agents.
And enterprises are gravitating to them. In a survey of more than 1,300 companies last year by LangChain, the developer of eponymous AI software framework and LangGraph orchestration tool, 78.1 percent said they plan to develop their own agent and put it into production. In addition, 51.1 percent said they already haven an agent in production, likely from vendors like Google, Salesforce, and Microsoft.
Foundation model-based AI agents are just emerging. The language they rely on is for communications, like nouns, verbs, and adverbs, Pandey wrote, and most AI-native assistants are created as simple agents using singular foundation models. However, that is changing fast.
AI models and foundation models are beginning to work with other languages, such as those of health and medical science (such as how proteins combine and fold), material science (how molecules come together to create materials), and embodied AI agents (f how physical actuators combine to create movement).
“These foundation models of all kinds are just now beginning to reason, plan, have memory and think about ambiguous objectives,” Pandey wrote. “These are hard problems yet to be solved for, but progress is being made quarterly.”
There are subject matter expert (SMEs) agents that can collaborate to solve large and more difficult problems, and there are more accurate and efficient because of the collaboration than foundation models like OpenAI’s GPT-4o or o1, according to Pandey. Also, vendors – including Cisco – are building multi-agent applications (MAAs) that combine foundation and machine learning models with tradition software for larger problems. The MAAs include such offerings as Cisco AI Assistant and Salesforce Agentforce.
“The next logical step in the evolution will be to create an ensemble or workflows of MAAs,” he wrote. “This Ensemble of MAAs would collaborate across vendor boundaries and across the Internet to solve for larger business (or personal or social or physical) workflows. An Ensemble of MAAs are the ‘internet’ version of agentic computation.”
For AI agents to thrive, the infrastructure must be done correctly, according to Pandey.
“By combining advanced computing capabilities with quantum-safe security, we can build infrastructure that not only enables the next wave of AI innovation but remains secure for decades to come,” he wrote. “The organizations that embrace this shift early will be best positioned to leverage the transformative potential of collaborative agentic software.”
That’s very interesting and novel — a good initative by Cisco’s Outshift in my view. It seems to match the 2025 vision and recommendations of HiPEAC (European High Performance — Edge And Cloud computing or Embedded Architecture and Compilation?) relative to “Orchestrating technologies for distributed agentic AI”, and, “Open protocols for distributed agentic AI systems” (pages 6 and 8 in: https://www.hipeac.net/vision/2025.pdf ). Developments to be followed in the coming years to be sure.