Agents · Open source
RFP Agent on Microsoft Foundry
This demo repository shows the progression from model-only reasoning to grounded enterprise agents. It starts with a single-agent assessment, adds Foundry IQ retrieval over prior opportunities and proposals, grounds an agent in CRM, architecture, and cost tools, and adds tracing and evaluation. Harness Office renders the orchestration as an observable office where a lead agent plans, delegates to a style-review agent, and asks for human approval before a separate agent may search the web.
- My role
- Creator and builder
- Started
- Updated
- Visibility
- Open Source
- Built with
- Microsoft Agent Framework · Microsoft Foundry · Foundry IQ · Azure AI Search · OpenTelemetry · Python
- progressive demos, from baseline to observability
- 5
- collaborating agents with distinct roles
- 3
- approval before any web research
- Human
Problem
Enterprise agent demos often jump straight to a polished answer, hiding the grounding, delegation, and approvals that make agents trustworthy.
Solution
A step-by-step set of demos where every retrieval, delegation, retry, and approval is a real tool or middleware event and is visible to the user.
Architecture
Synthetic opportunities and proposals are indexed into Azure AI Search knowledge bases used through Foundry IQ. Agent Framework agents call bounded tools, delegate to sub-agents, and publish through a review gate, with traces exported to Foundry, OTLP, or Application Insights.
Lessons learned
Making agent work visible changes the conversation with stakeholders. People trust an agent more when they can watch it plan, cite sources, and wait for approval.
More screenshots (2)
agents · rag · human-in-the-loop · observability