An agent in three lines — and what happens next

Microsoft demonstrated an agent built on Microsoft Agent Framework that runs as a Foundry Hosted Agent with three lines of C#—without a Dockerfile, web server, manual authorization, or scaling configuration. The same blog series also covered bringing a personal finance agent to production, and the count there is different: at least four separate areas of engineering work. The gap between these two articles shows where businesses are losing money on AI agents in 2026.

Microsoft has also released channels for agents and workflows: the same agent can be exposed as an OpenAI Responses-compatible service, connected to Telegram, handed off to another agent via A2A, or served through an MCP client. A compact shared hosting core plus modular integrations for each channel turns the communication protocol into configuration. For integrators, this removes a common headache: instead of writing a separate adapter for each channel, they now change the package.

If the agent is published as a Native AOT application—the standard path for cold starts in the cloud—ahead-of-time compilation removes unused code and requires reflection alternatives. In this situation, managed tests do not test what will reach production: a green CI run says nothing about the compiled binary's behavior. MSTest 4.4 closes the gap by generating the test list at build time, before trimming, without rewriting test classes.

A small detail, but details like this are what separate “works for us” from “works for the client.”

For an agent, TTU is the time until the business can trust it to take an action without an engineer watching. The first console response is unrelated to this milestone: three lines of code produce it almost immediately, while trust requires observability, confirmations, memory, honest tests, and the right channels.

In KT.Team projects, this means MCP as a unified protocol for tool access; LLM & Security Gateway as a confirmation and control layer before an agent running on Claude or GigaChat writes anything to a live system; and RAG with versioned memory instead of a chat that forgets context the next day. The seemingly simple result—an agent you can trust—requires exactly this hidden work.