From Rule-Based Bots to Virtual Teammates

▶ Watch (01:00)

Teams launched chatbot support in 2017. Those rule-based bots still have 50 million monthly users, but LLMs changed the model in 2022: agents could infer intent from plain language and act on it. Microsoft’s February Work Trend Index found 80% of leaders expect employees to delegate tasks to AI the way they would a team member. Gino’s team proved the model internally with Athena, a DevOps agent built at a hackathon last fall. In the Athena bug-fix video demo (07:18), it tracks PRs and build status in chat. Engineers using it close PRs 52% faster and hit 60% fewer build failures.

Building a Teams Agent in Minutes with AI Library 2.0

▶ Watch (11:37)

Teams AI Library 2.0 ships with a CLI. One command installs it; a second scaffolds a typed project from a template. The echo template produces a working agent in under ten lines of TypeScript.

“You can build an agent in just a couple lines of code now.” — Gino Buzzelli

From there, VS Code’s dev tools bring the agent live for local testing in the VS Code echo agent scaffold demo (12:14). The on-message handler receives input, and the developer adds whatever LLM reasoning they need. The manifest folder has everything required to publish to the Teams store.

UX Layer: Conversation Surfaces, Adaptive Cards, and Mobile

▶ Watch (15:36)

Building on Teams gives an agent the conversation surface for free: direct messages, group chats, and a meeting sidebar pane (in preview). Developers focus on business logic, not the conversational interface. The Relic Cloud backlog demo shows the pattern: at-mention the agent in a group chat to check sprint status, move to a private one-on-one to find who has capacity, then add it to a standup to update the team.

Adaptive cards handle rich output. Charts and video now work inside cards, as do citations with drilldown overlays. All of it runs on mobile without extra code.

Skills and Knowledge: MCP, Agent-to-Agent, and Three New APIs

▶ Watch (22:55)

MCP lands in Teams AI Library with one line of code. Point the agent at an MCP server and it inherits all the skills that server exposes; add more later without republishing. Agent-to-agent protocol, announced in the Build keynote, lets a Teams agent delegate to any other agent supporting the standard, regardless of where it runs. Three knowledge APIs entered preview: Meeting AI Insights, Copilot retrieval, and a persistent memory module. Vikram wired an MCP client to a Composio LinkedIn server in the VS Code MCP client + LinkedIn live post demo (40:47) and posted to LinkedIn live.

▶ Watch (46:47)

Three partners showed production agents built on the Teams platform. LexisNexis Protégé joins a Teams meeting like a subject-matter colleague, identifies non-buyer clauses in an M&A agreement, and redrafts indemnity language against the firm’s own document management system. Zoho’s project management agent runs in a private meeting sidebar for back-channel queries, then switches to the public meeting chat when the user is ready to share findings. Kubiya’s DevOps agent diagnosed a checkout-service 500-error outage in a Teams on-call channel: traced the bad commit on GitHub and completed the revert and redeployment without anyone leaving Teams.

Notable Quotes

we’ve been seeing 35% growth month over month of adoption Gino Buzzelli · ▶ 9:46

engineers using Athena today are closing their PRs 52% faster Gino Buzzelli · ▶ 10:18

You can build an agent in just a couple lines of code now. Gino Buzzelli · ▶ 11:43

Key Takeaways

  • Athena, a hackathon-built DevOps agent, cut PR close time 52% and build failures 60% for 5,000 Microsoft engineers.
  • Teams AI Library 2.0 scaffolds a functional agent in TypeScript in under ten lines of code with one CLI command.
  • MCP support, agent-to-agent protocol, Meeting AI Insights API, and Copilot retrieval shipped or entered preview at Build 2025.