Taking agentic AI to the next level, TM Forum’s Innovation Hub develops architecture to enable agent-to-agent interaction between different organizations.
Vodafone, Google Cloud, TM Forum demo inter-organization AI agent collaboration
A new project at TM Forum’s Innovation Hub has taken agentic AI to the next level by enabling agents from different organisations to interact with each other. Demonstrated at DTW25-Ignite, the production grade reference implementation is led by Vodafone and Google Cloud.
Last year, TM Forum members collaborated within the Innovation Hub to develop a generative AI search tool called AIVA that is integrated into TM Forum’s website. The AIVA architecture has since evolved into an “Agentic Architecture”, and the agent makes it easier for members to find relevant material and specifications. Further work has seen AIVA built into the ODA Canvas so that it runs as a modular, composable component with API-first capability.
The next step for AIVA was to allow it to talk to an agent from another organization, an advance that Vodafone’s Head of New Technologies and Innovation, Lester Thomas, demonstrated live.
“We're actually implementing agent-to-agent interaction across different organizations,” said Thomas.
The demonstration showed AIVA talking to Vodafone’s AI for Enterprise Architects (AI4EA), which the operator built to support its enterprise architecture team. It is a multi-agent architecture, comprising a supervisor agent and specialized agents that perform tasks such as strategic insights, gap analysis and diagram creation.
The Vodafone agent has been linked to TM Forum’s AIVA agent running on ODA Canvas via model context protocol (MCP). Google Cloud’s Agent2Agent (A2A) protocol enables interaction between the agents.
“As an enterprise architect, if you ask our agent a question and it involves some standards, it can go and get advice from the TM Forum by talking to the [AIVA] agent. The TM Forum agent is also a multi-agent architecture, and it can ask different sub-agents about specific standards in the Forum,” said Thomas.
For example, if there is a question about APIs, the agents would consult multiple sub-agents and come back with information on which TM Forum APIs should be used for a particular use case.
“It’s the future way of how you get all the great work of the TM Forum distilled into relevant knowledge for employees in Vodafone…This is a genuinely useful tool for enterprise architects,” he added.
Krishnamurthy Srinivasan, Head of Data and AI/ML Solutions at Google Cloud, said the new inter-organizational interaction between these agents can “significantly enhance the productivity of the enterprise architects and developers within Vodafone.”
He also explained how the MCP and A2A protocols are complementary in enabling inter-organisational agent-to-agent collaboration. “MCP is used for an agent to be able to access any of the tools a standard way, and A2A is the protocol for the agents to interact with each other,” he explained.
Standards approach opens way to more functionality
Because the architecture is built on open standards in an API-first, modular approach, any tool that supports MCP can interact with the agents. For example, the demo showed automated code generation based on Vodafone requirements and TM Forum API specs.
Thomas noted how Vodafone software engineers can benefit from such agent-to-agent collaboration, as their own coding assistant can consult with TM Forum’s AIVA and get the “absolutely correct” standards to apply.
“What the coding assistant generates will natively adopt the standards… Imagine what it’s like for a new software engineer in Vodafone. They might have come from a different industry and never even have heard of the TM Forum. They get the Vodafone standard development environment, they start building, and they're already adopting all the TM Forum standards with no effort whatsoever,” he said.
For Vodafone, the inter-organizational agent-to-agent communication demo offers a glimpse of its target enterprise architecture, which eventually will allow internal agents to talk to external agents.
Thomas outlined that the target architecture vision across every function in Vodafone is to have domains with a supervisor agent and multiple sub-agents in each domain. The domain agents can talk to each other and even talk to other organizations, as this use case has proved, he explained.
“We think this will become normal. You will have agents in your organizations, and those agents will be empowered to talk to other agents in other organizations about optimizing your business,” he said.
Agentic AI milestone
Aniket Mhala, Head of Innovation Hub, TM Forum, said the inter-organization AI agent collaboration project as a pioneering milestone in cross-enterprise agent integration.
“This Innovation Hub project sets a good precedent for scalable, multi-organization AI agent ecosystems, demonstrating seamless interoperability and laying the groundwork for future enterprise-wide agentic frameworks,” he said.
In addition to agents’ integration, the project also successfully implemented “agentic observability”, enabling real-time failure detection, behavioral analysis, audit trails, and performance evaluation, which are critical for operational reliability.
“What the coding assistant generates will natively adopt ehthe standards… Imagine what it’s like for a new software engineer in Vodafone. They might have come from a different industry and never even have heard of the TM Forum. They get the Vodafone standard development en, they start building, and they're already adopting all the TM Forum standards with no effort whatsoever,” he said.