AI
Why Agentic Automation Is Different This Time
Reasoning models are roughly seven months old. Everything else in workflow automation (forms, conditional logic, RPA, low-code) has been around for years. The problem those older tools share: when a business process changes, every script breaks and someone fixes it by hand. Previous technology couldn’t adapt as processes evolved.
Agents can now reason, act, and learn. That combination unlocks tasks that couldn’t be automated before. Bolaris’s starting advice: don’t invent a new use case. Take a process you already have and make it better with an agent.
Building an Agent: Code, Knowledge, and Actions
An agent needs three things: a model, instructions, and tools. In the VS Code single-agent build and ADO ticket creation (13:51), Bolaris wrote those as constants, called the SDK, and hit run. The agent appeared in Azure AI Foundry. For knowledge, options include Bing custom search (scoped to specific sites, now in public preview), Azure AI Search with multi-query agentic retrieval, SharePoint, and Microsoft Fabric. For actions, any OpenAPI v3 spec becomes callable. Azure Logic Apps integration, which also launched that day, let the agent create ADO issues and send email directly.
Security, Governance, and Observability
Agents making real-world calls need access controls. Entra ID integration (coming soon) gives each agent a scoped identity with specific permissions, the same model Entra uses for applications. Prompt shields block direct manipulation and catch indirect attacks, where malicious web content tries to hijack agent behavior. A Red Teaming Agent sends adversarial inputs across violence, harassment, and sexual content categories before the agent ships to production.
Every interaction is stored as a thread; each agent turn is a run. The trace view shows token usage, latency, and evaluations per run. Intent resolution is on by default; safety scores are optional.
Multi-Agent Orchestration and Human-in-the-Loop
One agent handling everything doesn’t scale. Bolaris built four: a triage agent fans emails to a FAQ agent and a RAG agent. When neither could answer when GA was launching (a question missing from the help docs), the orchestrator escalated. In the live multi-agent pipeline with M365 human escalation (38:16), the system messaged Bolaris in M365 Copilot. He typed “Today”. The workflow resumed and sent the customer a full reply seconds later.
“These are self-healing.” — Mads Bolaris
The Semantic Kernel Process Framework defines this flow with events. Execution order stays consistent; connected agents vary their path on every run.
Protocol Interop: MCP, A2A, and Copilot Studio
Azure AI Agent Service now supports three protocols. MCP adds any MCP server’s tools to an agent; Semantic Kernel code added GitHub’s MCP server in a few lines, and the agent queried pull request activity live. A2A (Google, announced early April) enables cross-vendor agent calls. SAP demonstrated this by routing tasks across Azure, their own, and Google agents in one portal. Copilot Studio integration lets a no-code worker add a Foundry-built agent as a sub-agent in their own pipeline. Microsoft is co-developing both protocols to add enterprise-grade capabilities.
Notable Quotes
You can now automate the unautomatable, Mads Bolaris · ▶ 6:53
what we’ve promised is AI agents can help. Mads Bolaris · ▶ 4:42
These are self-healing. Mads Bolaris · ▶ 42:08
Key Takeaways
- Azure AI Foundry Agent Service reached GA at Microsoft Build 2025, with 10,000 customers and millions of agents created in three months of public preview.
- Semantic Kernel Process Framework lets developers declaratively define multi-agent workflows with consistent step order, unlike connected agents that vary execution path.
- Prompt shields now block indirect injection attacks where malicious web content tries to hijack agent actions, and Entra ID integration is coming to give agents scoped permissions.