Cognizant Open-Sources Its Neuro AI Multi-Agent Accelerator

‘[We] want to establish this as the de facto standard by which people create these types of scaled multi-agentic systems. There are other platforms out there for creating multi-agent systems. We found none of them to be at the same scale with the same level of interoperability, the simplicity to build these systems, and the safety features that we have. So we decided to put this out there in the service of the AI and developer community and encourage everyone to use it,’ says Babak Hodjat, Cognizant’s chief technology officer of AI.

Global solution provider Cognizant Thursday said it has open-sourced its Neuro AI Multi-Agent Accelerator technology for prototyping and building agent networks across any industry vertical.

The Neuro AI Multi-Agent Accelerator is the platform and framework that Cognizant uses to build multi-agentic systems safely and at scale, said Babak Hodjat, chief technology officer of AI at Teaneck, N.J.-based Cognizant, which is No. 8 on CRN’s Solution Provider 500.

Cognizant uses its Neuro AI Multi-Agent Accelerator to create multi-agent systems for internal use and for its clients, but is now bringing it to open source to promote the creation of multi-agentic systems, Hodjat told CRN.

“This is going to be an academic research license, obviously,” he said. “But we also want to establish this as the de facto standard by which people create these types of scaled multi-agentic systems. There are other platforms out there for creating multi-agent systems. We found none of them to be at the same scale with the same level of interoperability, the simplicity to build these systems, and the safety features that we have. So we decided to put this out there in the service of the AI and developer community and encourage everyone to use it, because I think people will benefit from being able to create agent networks that then they can connect to other agents.”

The repository for Cognizant’s multi-agent accelerator will be located on GitHub and known as neuro-san, where “san” is short for system of agent networks, Hodjat said. There will also be a studio available, neuro-san-studio, with a number of pre-built templates, he said.

“Those pre-built templates include, and maybe this will amaze you, an agent network that can create agent networks,” he said. “It’s a network of agents. You describe your process or your company, or the task at hand, and it will automatically design and create an agent network for that particular task. Obviously, there’s work to be done to ground these systems, but it’s a great starting point to explore.”

The academic research license gives developers and researchers the opportunity to experiment with the Neuro AI Multi-Agent Accelerator, as well as contribute back to the technology, Hodjat said. However, that license does not allow commercial use, he said.

“I think we need the libraries to be in the hand of professionals who can actually build them out and ground them for the enterprises,” he said. “And we think we’re the best people to do so for that. People have to come see us, but if you want to kick the tires, if you want to contribute as part of the community to the code base and help extend it and expand it and make your own agents interoperable with it, this can end up being the fabric that allows this interoperability with other agents.”

Hodjat said the system is already interoperable with Google’s Agentspace and Salesforce’s Agentforce, and supports Anthropic’s MCP (model context protocol).

“Many companies, including Google and OpenAI and so forth have signed up to the agentic worldview,” he said. “It also supports A2A, or Agent2Agent, announced by Google recently. It’s been designed to be fully interoperable. As a developer, as someone that’s working in academia or research, you have full access, and you can play around with it. It’s just the commercialization aspect that is restricted.”

The Neuro AI Multi-Agent Accelerator allows users to run agent networks and serve them, Hodjat said. Users can have multiple servers, with each server running its own agent network, and agents from one server can talk with agents to another server.

“It allows that kind of complexity in a very straightforward, simple way, actually,” he said.

Hodjat said there are other open source multi-agent accelerators available, including an open source library called LangChain on which the Cognizant system was built.

“LangChain is almost a de facto standard now for agentic systems, and it allows you to build agentic systems that are LLM-agnostic,” he said. “So if you want to use Google Gemini or Open AI GPT-4 or Anthropic or what have you, or open source networks like Llama, you can do that very easily. We didn’t reinvent the wheel on that. “

Other open source multi-agent accelerators include Microsoft AutoGen and LangGraph, he said.

Cognizant is “client zero” for its Neuro AI Multi-Agent Accelerator, Hodjat said.

“We’ve used it and it’s in deployment now for our own intranet,” he said. “And what’s amazing is that the intranet has many agents in it. This would be the biggest modern multi-agentic system you’ve seen. It includes HR, finance, legal, and sales. Just imagine the scale and the security aspects and the safety aspects that have to go into identifying your intranet, as well as the fact that you need to set it up in a way that is incrementally extensible. And that’s one of the things that this system does very, very well.”

Cognizant already has a number of commercial clients for its Neuro AI Multi-Agent Accelerator, including Telstra, the Australian telecommunications company.