Microsoft Unveils New AI Agents in Semantic Kernel, Adding ChatCompletion, Group Chat,

Microsoft’s Semantic Kernel now ships with a full‑featured agent framework that pushes the platform beyond simple chat completions into autonomous, goal‑driven workflows. The new ChatCompletionAgent is a single‑agent wrapper that couples a conversational model—currently Azure OpenAI’s gpt‑4o—with plug‑in tools and stateful memory, allowing it to pull data, invoke functions, and keep context across turns. In a code sample posted by Brian Spann on March 1, the agent is wired to an “OrderPlugin,” “CustomerPlugin,” and “RefundPlugin,” then instructed to verify a user’s identity, look up order status, and process a refund if the shipment is overdue. The agent’s response stream demonstrates the “tool‑use” loop: it calls `get_order_status`, evaluates the delay, triggers `process_refund`, and finally returns a confirmation, all without human intervention.

Beyond single agents, Microsoft introduces AgentGroupChat, a multi‑agent orchestration layer that lets several specialized bots collaborate on a shared task. The framework handles message routing, turn‑taking, and shared memory, effectively turning a collection of agents into a coordinated team. VentureBeat notes that this replaces Microsoft’s earlier AutoGen project, consolidating “observability, governance, and lifecycle management” under a single API surface. The shift signals a strategic move to give enterprises a more manageable way to deploy complex AI pipelines, where a sales‑assistant, a compliance checker, and a logistics bot can each contribute their expertise while the group chat maintains a unified dialogue.

The orchestration tools also expose function‑choice behavior and execution settings that let developers fine‑tune how aggressively an agent should invoke external tools. In Spann’s example, the `AzureOpenAIPromptExecutionSettings` are configured with `FunctionChoiceBehavior.Auto()`, enabling the model to decide on‑the‑fly whether a function call is needed. This mirrors the “reasoning → tool use → memory” loop Microsoft describes as the core of an agent’s capabilities. By exposing these knobs, the Semantic Kernel gives developers granular control over latency, cost, and safety—critical considerations for enterprise deployments that must balance responsiveness with compliance.

The broader Azure AI rollout at Ignite 2024, highlighted by Forbes, positions the agent framework as one of the “Top 5 Azure AI Announcements.” According to the outlet, the new agents are designed to integrate tightly with Azure’s existing ecosystem—Azure OpenAI, Azure Functions, and Azure Cognitive Search—so that enterprises can spin up end‑to‑end solutions without stitching together disparate services. The Register adds that Microsoft is publicly previewing “autonomous Copilot AI agents” that will eventually run on this same infrastructure, hinting at a future where Copilot‑style assistants can not only answer questions but also execute business processes autonomously.

Analysts see the move as Microsoft’s answer to the growing demand for “AI‑as‑a‑service” that does more than surface information. By embedding tool use and memory directly into the agent abstraction, Microsoft hopes to lower the barrier for developers to build production‑grade assistants that can handle refunds, schedule meetings, or even orchestrate multi‑step data pipelines. The framework’s open‑source roots—Semantic Kernel is a .NET library—also mean that the community can extend it with custom plugins, potentially accelerating adoption across sectors that have been slower to embrace large language models.

In practice, the new agents could reshape how companies staff their support desks and internal workflows. A single ChatCompletionAgent, armed with domain‑specific plugins, can triage tickets, pull CRM records, and issue refunds—all while logging each interaction in a persistent chat history. When multiple agents join forces via AgentGroupChat, the system can delegate complex requests to the most suitable specialist, reducing hand‑off friction. As Microsoft phases out AutoGen in favor of this unified framework, the company is betting that tighter governance, built‑in observability, and seamless Azure integration will give it the edge in the race to deliver truly autonomous enterprise AI.