Chandra Chivukula, VP Engineering, Microsoft, On Microsoft Teams as an Agentic Platform

Prefer audio?

Who should listen: M365 and Teams service owners, collaboration and information management leads, and anyone planning how agents like Facilitator and Channel Agent will fit into the way their organisation works in Teams.

Guest: Chandra Chivukula — VP Engineering, Microsoft Teams

OpenClaw showed how much people like working with agents through their messaging app, and Teams is where most enterprises already message. Chandra Chivukula, who leads engineering on the core Teams clients and infrastructure, joins Tom to talk about Teams as a platform for both humans and agents. He covers Microsoft Scout, Facilitator, Channel Agent and Copilot in group chats, how his own engineering teams are working with coding agents, and why moving work from group chats into channels makes the AI more useful for everyone.

Many thanks to AVI-SPL, sponsor of this episode.

Key insights

• Tom's starting point: OpenClaw's take-off showed that people like talking to agents through their messaging platform. With roughly 350 million monthly users, Teams is the enterprise version of that surface. ▶ 1:33
• Chandra says Microsoft has to bring agents to its mission-critical customers without losing the security and compliance they expect from Teams. Microsoft Scout, launched at Build, is the example: built on the OpenClaw agent framework by a team under Omar Shahine, it runs inside the Microsoft 365 compliance boundary with access to your corporate data. For now it is a separate desktop app, available through the Frontier program. ▶ 2:58
• In Chandra's view it is very hard to "surgically implant" agents into the way people work and expect much to change. How you structure your teams, chats and channels is an investment decision that sets how much use agents can make of them. ▶ 5:57
• His engineering teams had both Claude Code and Copilot CLI, and are now all in on Copilot CLI, mainly because of the range of models behind it: Anthropic, OpenAI, Microsoft's new MAI coding model and others. Tom does the same as a non-developer, having one vendor's models build something and another's check it. ▶ 6:57
• The people getting the most from AI are not defined by seniority. Principal engineers who were already excellent are seeing 5x to 10x productivity. Engineers one or two years out of college are researching where each model is strong and weak and mixing models to suit. What counts, Chandra says, is adaptability, and redoing that research every six weeks or so as new models land. ▶ 8:29
• Building the team's own harness took a while: its own knowledge base and self-running loops, so that every bug fix and feature adds to a corpus that makes the next one easier and stops the same mistakes recurring. Chandra says this has visibly raised the quality of their output over the past four months or so. ▶ 11:29
• The same pattern carries over to meetings. You invoke Facilitator mid-meeting to go and check on something, and a minute or two later it comes back with a short report. Tom's wow moment was asking Facilitator in voice mode to write a PBI on a backlog call. It wrote one and posted it in Loop, though it could not yet push it into the backlog. ▶ 12:28
• Chivukula draws the line between chats and channels. Group chats suit short-lived, specific work. Channels are a longer-running construct where people come and go and work on shared files in the SharePoint mapped to the channel, with its governance. That is why a channel was the logical first home for an agent. Channel Agent is in preview. ▶ 16:28
• You can now add Copilot to a group chat as a kind of digital worker. Tom thinks no platform has nailed multiplayer AI yet, Microsoft included, and sees a big unlock if it does. Chandra wants Copilot to feel like a virtual coworker of limited capability that does small things reliably, rather than something you go to a separate app to use. ▶ 17:31
• Discoverability is the case for channels. Group chat messages cannot be searched by anyone outside the chat, but public channels are indexed in Microsoft Search, so other teams can find and reuse what a project has learnt. Private channels can be opened up once a sensitive project has launched. ▶ 21:20
• Chandra's example is a scheduled task in Microsoft 365 Copilot Cowork that searches each morning for anyone else working on a similar initiative. It will find nothing if the whole company works in group chats. Channels give an accruing value to the company knowledge base, not a point-in-time one. Tom adds the selfish reason: the channel agent or Facilitator can answer any question about the project across its whole history. ▶ 26:19
• On mobile, Catch Up gives a card-based, one-handed triage of @mentions and unread messages rather than the whole activity feed. Chivukula hopes it makes catching up about five times faster. Tom says it is what pulls him back into his channels when travelling. Features for saved messages and VIP users are also rolling out, and Chandra would not be drawn on future features. ▶ 28:56

Insights summarised by AI from the episode transcript, reviewed by the Empowering.Cloud team.

Listen: Apple Podcasts · Spotify · Other platforms