AI agents bring with them the promise of being able to autonomously solve complex tasks put before them, from finding and analyzing the necessary data, choosing tools, and making decisions without human intervention to learning from their mistakes and adapting to changes. They also can collaborate with each other to coordinate their actions.
However, to do that, they need to be able to reliably and securely communicate with each other, which isn’t easy given that they can be created from disparate large language models and foundation models from different companies. As with other moments in IT history that led to a wide distribution and decentralization of computing – think the internet and cloud computing – protocols are needed to ensure that systems and software can communicate with each other.
That gave rise to such protocols as TCI/IP, HTTP, Voice over IP (VoIP), and Message Queuing Telemetry Transport (MQTT), all of which helped drive communication and interoperation.
Protocols for AI agents are now cropping up, such as the Model Context Protocol (MCP) from Anthropic, allowing LLMs to access external data sources and tools, and Agent2Agent (A2A), announced by Google and more than 50 other tech vendors and consultants in April to make it possible for agents to work together.
“Enabling agents to interoperate with each other, even if they were built by different vendors or in a different framework, will increase autonomy and multiply productivity gains, while lowering long-term costs,” Google said at the time.
The Need For Protocols
A2A came a month after Cisco Systems’ Outshift incubation arm – along with LangChain, which developed an open framework for developing AI applications, and Galileo, which has a AI-powered text-to-UI design tool – introduced AGNTCY, an open source protocol similar to A2A to enable organizations to find the best agents for job, bring them into workflows across any framework or vendor, run these multi-agent systems both security and at scale, and monitor and evaluate their performance, essentially creating what Vijoy Pandey, senior vice president of Outshift by Cisco, calls the “internet of agents.”
Google last month donated the A2A protocol to The Linux Foundation, and this week Cisco and its now 70-plus tech partners are doing the same with AGNTCY, which will be a new project within the nonprofit organization. A protocol like AGNTCY shouldn’t be controlled by Cisco or any single vendor, Pandey tells The Next Platform. It needs to be managed by a neutral entity.
He also wanted the governance of AGNTCY “to be reflective of some heavyweight organizations,” which now includes Dell, Google Cloud, Oracle, and Red Hat, giving it credence in such areas as hyperscale, open source, and infrastructure.
Cisco is positioning itself as the foundational infrastructure layer for AI, starting with its Silicon One networking chip and moving up the stack. AGNTCY is reflective of that, and with such wide industry support will be an important tool given the decentralized nature of agentic AI and the need for these agents – which often are built for particular tasks, from industry expertise like coding, customer service, financial analysis, and sales and marketing, and jobs, such as perception, reasoning and planning, and learning – to pull together their particular skills.
“We’re moving from deterministic computing to probabilistic computing,” Pandey says. “The state-of-the-art in software used to be cloud-native computing –cloud computing – and we’re building an abstraction layer above that. This is agentic computing, or agentic-native computing. … Embodied agents, scientific agents, social agents, they all need to collaborate to solve large problems and they need to collaborate across a lot of diversity, a lot of heterogeneity. All these agents will be like subject matter experts and, like a team, they all need to come together for a larger goal.”
One Platform To Unite Them All
The aim is to create a platform where they can all come together, communicate, and collaborate.
“They will be coming from different vendors,” he says. “They will have used different development frameworks to get created, they will be running on different clouds – Amazon, Google, on-prem – they will have belonged to different organizations. They will take on different profiles and personas as they run in their environment because they in some ways mimic humans.”
AGNTCY which includes a identity and discovery platform and collaboration and communications frameworks, uses the principles of peer-to-peer (P2P) communications that emerged from the trend toward decentralized computing, he says. A2A and MCP servers are discoverable through AGTNCY’s agent directories, with transparency coming via its observable SDKs. AGTNCY also includes a message transport over the Secure Low Latency Interactive Messaging (SLIM) protocol.
The agent discovery capability relies on the Open Agent Schema Framework (OASF) that allows agents to discover other agents and understand their capabilities, while the messaging function includes cryptographically verifiable identity and access control to ensure security even when working across organizational boundaries.
Security Is A Priority
Security will be a key part of AGNTCY, as it is with all parts of AI. Agents will be grabbing corporate and sensitive data to do their work, which potentially exposes that information to outside threats. Security questions are swirling around MCP servers, with Trend Micro researchers earlier this month saying they found 492 MCP servers with no client authentication or traffic encryption. In all, these MCP servers offer access to 1,402 MCP tools, they wrote, with more than 90 percent of them providing direct read access to the data source. The data is available to anyone using natural language tools, which means they don’t have to know how to code to steal it.
Pandey acknowledged the need for hardened protections, adding that in agent-to-agent communication, when doing something as necessary as deciding what a task is, an escalation of privileges is necessary. In agentic AI, authorization and access control not time-based, which is done with humans, or service-based, as with machines and software.
“It’s task-based,” Pandey says. “The hooks are there for it to allow task-based escalation of privileges and then bring people back to base level. If you’re writing agents in the right way, if you’re writing MCP servers in the right way, you can use this task-based authorization and actually build a pretty secure agent on a pretty secure MCP server.”
It’s going to be needed, because MCP servers are supported in most parts of AGNTCY, including in the Agent Directory, the schema framework, the identity platform, and the SLIM messaging layer.
That said, there is a challenge Cisco and the vendors behind AGNTCY are working on.
“Let’s say you don’t incorporate that task-level capability in your MCP server or in your agent,” he says. “How do you then handle it on a separate plane? That’s something that we’re working that you’ll see us coming out with in the future.”
Into The Future
Even before donating AGTNCY to The Linux Foundation, Outshift and partners were adding to it. Earlier this month, the group introduced coffeeAGNTCY, a reference application allowing organizations to experiment building multi-agent software with MCP, A2A, and AGNTCY components. At the same time, they released the AGNTCY Application SDK to help with integrating such components as SLIM initially, with others like Observe-SDK and Identity coming later.
Going forward, Pandey said work is continuing on building applications on AGNTCY, with Cisco already creating a few proprietary and open source tools, including one for AI engineers that is being leverage by a host of site, reliability, and engineering (SRE) teams that are part of the Cloud-Native Operating Excellence (CNOE) community. There also is work on networking and security pipelines being done with telcos and large enterprises.
“What you can expect to see in a year is a lot more companies leveraging agency to build real-world enterprise applications,” Pandey says. “That’s the big missing piece. … We need to make these things real-time and interactive.”
He also expects the number of industry contributors to AGNTCY to grow to more than a hundred.
Hell Freezes Over: Cisco And Nvidia Cross-Pollenate AI Networking
UPDATED Networking giant Cisco Systems and AI platform provider Nvidia have hammered out a deal to mix and match each other’s technologies to create a broader set of AI networking options for their respective and – importantly, prospective – customers. Nvidia has been a competitor of Cisco’s in the datacenter …
AI Pervades The Cisco Stack, But Is Only Starting To Drive Sales
Like many other suppliers of hardware and systems software, Cisco Systems is trying to figure out how to make money on the AI revolution. Thus far, like its OEM and ODM peers, Cisco has had some success building up an AI pipeline, but it is measured in single digit billions …
Cisco Pulls Together A Quantum Network Architecture
For decades, discussions around quantum computing has felt similar to family driving vacations, with someone in the back seat constantly asking “are we there yet?” And like those long drives, the answer usually has been “not yet, but soon,” with “soon” doing a lot of the work in the reply. …
Be the first to comment