Last month saw Microsoft update its Azure AI Foundry with agentic capabilities, getting a little help from both Cisco and the open-source community.

Through the new Microsoft Agent Framework, Azure’s agent factory touts an open-source software kit for devs to get to grips with all things agentic AI. This includes universal API integrations as well as agentic communication through Google’s Agent2Agent (A2A) standard.

The Cisco partnership came through enhanced tracing and telemetry for multi-agent systems, specifically through Outshift by Cisco, the networking giant’s incubation engine division, and the OpenTelemetry (OTel) framework.

“Microsoft had an idea to update OTel for multi-agent tracing and they brought it into the discussion [with the OTel community]," revealed Yina Arenas, CVP of product for Azure AI Foundry at Microsoft. "In the discussion it was found that Cisco OutShift team had similar ideas, so both Microsoft and the Cisco OutShift team partnered to update the OTel standard for multi-agent tracing.”

These enhancements involved greater visibility into agent workflows and tool call invocations, which capture and analyze each usage of tools during a session. Ultimately, both attributes aid with agentic debugging, optimization, and compliance.

As a result, Azure AI Foundry can claim to offer a unified observability solution that monitors and traces agents built on diverse frameworks, including Microsoft Agent Framework, LangChain, LangGraph, and the Agents SDK from ChatGPT behemoth OpenAI.

In one case study, consultancy giant KPMG used the framework for its cloud-based Clara AI auditing platform, building a multi-agent design fabric capable of orchestrating agents into coordinated workflows with minimal context switching across tools and platforms.

But while Azure devs stand to gain from the OTel enhancements, both Microsoft and Outshift pointed to a universal benefit in their attempt to tame the increasingly unruly landscape of AI agents.

“It’s about improving the interoperable format for everyone using it. Historically, our products like AppDynamics and Cloud Observability integrate with Azure, but this collaboration is about evolving the open-source framework itself,” an Outshift representative explained to SDxCentral via email.

The unit has some form here, having developed an agentic directory named Agntcy, which it donated to the Linux Foundation earlier this year.

When interviewed by SDxCentral regarding the Linux news, Vijoy Pandey, GM and SVP of Outshift by Cisco, explained that with numerous agents already deployed, “the world of eight billion people will feel more like 80 billion from a bandwidth perspective.”

“In order to build these collaborative systems, agents need to be able to find each other, verify their identities, and share context without expensive custom integration work,” Pandey commented, highlighting Agntcy’s capability of exchanging messages efficiently and securely at scale through secure low-latency interactive messaging.

The OTel collaboration works in the same vein, with Outshift focused on achieving enterprise-grade multi-agent observability by tackling non-determinism of agentic models, the challenge of integrating agents with siloed enterprise data, as well as the high costs of sustained GPU compute.

“Furthermore, because these agents are autonomous and access sensitive data, providing a clear, auditable trail for governance is a foundational security requirement. That traceability is fundamental to building the trust and transparency required for enterprise AI,” explained Outshift's representative.

“This OpenTelemetry work is essential. What the community did at the agent and multi-agent system level is the foundation for the observability components we are building. These fit into Agntcy’s four key pillars: discovery, identity, messaging, and observability,” they added.

MCP: the poster model for agentic AI

Also pivotal to the agentic piece is integration of model context protocol (MCP), the increasingly influential integration standard for AI agents which AI firm Anthropic released last year.

In Azure AI Foundry, Microsoft operationalizes MCP for enterprise use, with stored credentials, agent identity (Agent ID), and role-based access controls that ensure each tool invocation is authenticated, logged, and policy-bound.

Azure's MCP integration is also enabling enterprise extensibility by allowing internal systems such as HR, CRM, or procurement to expose MCP endpoints for secure agent access.

This extends to advanced ecosystem interoperability, with agents from different frameworks like Semantic Kernel and AutoGen able to communicate through shared MCP connectors.

Arenas identified early use cases such as unifying developer tools through a single interface encompassing APIs, embeddings, and vector stores across vendors.

The exec also believes that MCP marks a turning point in how AI agents connect to the broader digital ecosystem.

“With MCP support built into the Microsoft Agent Framework and Azure AI Foundry, developers can now connect any compliant tool or service, whether it’s a database, SaaS application, or enterprise system, without custom integration work.”

In addition, Arenas revealed the tech giant will continue to partner with Anthropic and the MCP community to update the standard and expand the ecosystem, as well as uphold its authorization spec and security best practices.

The CVP also answered the hazy question of how exactly networking teams use MCP, something SDxCentral explored in a recent article asking whether future networks are in need of adopting the standard.

“Cloud and networking teams use MCP to automate provisioning, routing, and monitoring through a unified protocol layer, while security teams leverage it for governed tool access, zero-trust enforcement, and full auditability,” said Arenas, who very much sees MCP as the plug layer of sorts for the AI economy.

“It allows reusable agentic skills and shared capabilities across open-source communities, enterprises, and even multi-cloud environments. The result is a consistent, compliant control surface for agent-to-tool interactions, combining openness with enterprise-grade trust.”

The Microsoft exec linked back to the Outshift/Cisco collaboration in delineating how MCP works with Azure AI Foundry’s observability functions. The former, she explained, standardizes how agents connect to external tools and systems, while the latter focuses on what happens inside and between agents.

In other words, MCP governs the interface layer, whereas Foundry’s agent-to-agent telemetry governs the behavioral layer, tracking reasoning steps, tool calls, and message exchanges across multi-agent workflows.

This is as Foundry Observability’s OTel-based agent-to-agent interaction builds an observability plane for debugging, compliance, and optimization, factors which aren't part of MCP’s basic objective.

“In other words, MCP governs the interface layer, whereas Foundry’s agent-to-agent telemetry governs the behavioral layer, tracking reasoning steps, tool calls, and message exchanges across multi-agent workflows,” Arenas explained.