AI
The Agent Store: A New Distribution Channel for Copilot Agents
Microsoft launched the Agent Store at Build 2025, a discovery hub inside Microsoft 365 Copilot where users find and install agents from partners, Microsoft, and independent developers. Install once and the agent appears across Teams, Outlook, and Office. The store works for users with or without a Copilot subscription.
“It’s the UI for your AI.” — Siffat Hingorani
For partners, the store offers reach to millions of Copilot users, accelerated time to market, and a Microsoft-vetted catalog. At launch, more than 70 agents were commercially available and the number was growing daily.
How to Build a Great Agent: Purpose, Prompts, and Instructions
Steven Abrahams distilled
“seven tips my team uses when working with agent builders.” — Steven Abrahams
Start with a clear purpose and a specific user scenario. Agents that try to do everything tend to do nothing well. Starter prompts are the first impression: six slots to show what the agent does with a specific data type or analysis, not just announce its name. For instructions, use positive language and test repeatedly.
“the ones that we’ve fine-tuned through testing.” — Steven Abrahams
API plugins must be defined in the OpenAI spec YAML manifest package before you can publish. For declarative agents, domain ownership verification is also required.
Making Your Agent Discoverable and Trusted in the Store
Your agent’s avatar, name, and short description are the first things users see in the store. Microsoft rejects listings with marketing claims (“best,” “number one”) or URLs. The short description should cover value in seven to nine words. Descriptions also feed Copilot’s search index, so keyword choice matters.
On feedback: custom engine agents can add thumbs-up/down buttons via the Teams AI library and retrieve the raw telemetry. Declarative agents get those ratings from the platform automatically. Build a review cadence and use the data to decide what knowledge sources or APIs to expand next.
Partner Demos: Miro and Learn365 Inside M365
Abrahams demonstrated two production agents already in the store. Miro (90 million users, customers including PepsiCo and Deloitte) built an agent that searches across a user’s boards by natural language. In the Miro board search live demo (28:31), a query for recent boards about Azure costs returned a rich adaptive card with a content preview, no board title or owner required.
The second demo, Zensai’s Learn365, generates training courses from inside Copilot Chat. A prompt asking for an AI basics course in Italian routed to the Learn365 API, confirmed the write operation with the user, and published the course to the Italian catalog, all without leaving M365.
Build and Publish a Declarative Agent in Minutes
Kiran Thomas built a working agent in under five minutes using the VS Code declarative agent build demo (37:20). The Microsoft 365 Agent Toolkit scaffolds the project with pre-filled JSON. One provision step deploys to the M365 tenant. Thomas added poem-response instructions and conversation starters, re-provisioned, and the agent went live in Copilot.
Publishing runs through Partner Center in three steps: enroll in the Microsoft 365 and Copilot program, create an offer, then submit the package. Microsoft reviews each submission and gives written feedback on failures. Compliance certification speeds up IT admin approval and unlocks usage analytics in the developer portal.
Notable Quotes
It’s the UI for your AI. Siffat Hingorani · ▶ 3:03
seven tips my team uses when working with agent builders. Steven Abrahams · ▶ 13:59
the ones that we’ve fine-tuned through testing. Steven Abrahams · ▶ 18:32
Key Takeaways
- The Agent Store launched at Build 2025 with over 70 commercially published agents and reach to a billion M365 users.
- Six starter prompts and tested instructions are the highest-leverage levers for agent quality and user adoption.
- Store submission takes three steps through Partner Center; Microsoft compliance certification accelerates IT admin approval.