Agents for Contoso

THE ENGAGEMENT

Two agents will star in the demo, and both answer a specific finding. The RFP responder drafts DDQ answers grounded in Contoso’s own answer library, with provenance on every line — that is finding D3, weeks of re-answering, addressed. The KYC chase agent fires autonomously when a KYC item goes stale and advances the onboarding state — that is D4, the email-based chase and the refresh backlog, addressed. Neither is a chatbot; each is a worker with a job the client named.

THE HARNESS

Copilot Studio in July 2026 makes both buildable without heroics. Autonomous agents and event triggers are generally available^⊙ , so the KYC chase can fire on a stale item rather than waiting to be asked. Multi-agent orchestration and agent-to-agent are GA^⊙ , so the two agents can hand off. MCP is the tool plumbing, GA over the connector infrastructure that inherits DLP^⊙ ; computer use is GA for the cases with no API^⊙ ; and Claude models are in the model list.^⊙

The discipline point: the Copilot Studio VS Code extension, GA since January 2026^⊙ , keeps these agent definitions in the same repo as the instruction stack and the data model. The demo agents are versioned artifacts, not clicks in a portal you cannot diff.

Provenance is the product

The RFP responder’s value is not that it drafts fast — anything drafts fast. It is that every drafted line carries a citation back to the answer-library source, so an SME reviewing the queue can trust or correct each claim in seconds. In a regulated firm, an answer you cannot trace is an answer you cannot send. Build the provenance first and the speed second; the reverse impresses no one who has to sign the DDQ.

In the field

Sketch the KYC chase agent’s trigger and its one autonomous action, and write down what it must never do without a human — advancing a stage is fine, sending a client-facing email may not be. That boundary is the difference between an autonomous agent and an unsupervised one.