Architecture Overview

High-level design of the AI-powered Software Request Approval system.

Table of Contents

  1. System Design
    1. Key Platforms
  2. Agent Topology
  3. End-to-End Flow
  4. Design Decisions
  5. Next Steps

System Design

Architecture Overview

The system uses a fan-out / fan-in multi-agent topology built on Microsoft Foundry Agent Service. A single Orchestrator Agent receives every software request, delegates parallel research to three specialist agents, aggregates the findings, and passes them to a Report Builder for final assembly.

Key Platforms

Capability Platform / Service
Agent orchestration & hosting Microsoft Foundry Agent Service
LLM reasoning Azure OpenAI gpt-4o (2024-11-20) or o3
Real-time web research Grounding with Bing Search (native Foundry connector)
Internal knowledge Azure AI Search (policy, vendor, report indexes)
Knowledge indexing Foundry Agent IQ (Knowledge Store)
Conversational front-end Microsoft Copilot Studio
Approval routing Work IQ + Power Automate + Teams Adaptive Cards
System-of-record updates Power Automate / Logic Apps
Monitoring Azure AI Foundry Evaluation + Application Insights

Agent Topology 

[Copilot Studio Bot] ──── intake form ────▶ [Orchestrator Agent]
                                                  │
                         ┌────────────────────────┼────────────────────────┐
                         ▼                        ▼                        ▼
              [Compliance Agent]        [Security Agent]        [Vendor Checklist Agent]
                         │                        │                        │
                         └────────────────────────┘────────────────────────┘
                                                  │  aggregated JSON
                                                  ▼
                                        [Report Builder Agent]
                                                  │  structured Markdown report
                                                  ▼
                                    [Power Automate → Teams Adaptive Card]
                                                  │  human decision
                                                  ▼
                                        [System Updater Agent]
                                   (ServiceNow / Entra ID / SharePoint)

All agent-to-agent communication uses the Connected Agents feature in Microsoft Foundry — no custom orchestration code required.

End-to-End Flow

End-to-End Flow Summary

Step Actor Action
1 Employee Submits request via Teams bot
2 Orchestrator Agent Parses metadata, fans out to research agents (parallel)
3 Compliance Agent Checks SOC 2, ISO 27001, NIST, FedRAMP, GDPR status
4 Security Agent Queries NVD/CVE databases, breach history, patch cadence
5 Vendor Checklist Agent Assesses vendor financials, restricted lists, support SLA
6 Report Builder Agent Aggregates findings into a structured Markdown report
7 Power Automate Routes report to approver via Teams Adaptive Card
8 Approver Reviews AI recommendation and decides in one click
9 System Updater Agent Writes decision to ServiceNow, Entra ID, SharePoint

Total elapsed time: ~3–7 minutes (vs. multiple days manually).

Design Decisions

Why no custom code? The entire core workflow runs as no-code/low-code using Foundry Agent Builder, Connected Agents, and Copilot Studio. Code is only needed for advanced customizations beyond the portal.

Why fan-out? Running the three research agents in parallel cuts total latency to the duration of the slowest single agent rather than the sum of all three.

o3 latency: If routing Compliance or Security agents to o3 for deeper reasoning, expect per-query times of several minutes. Configure these agents for async/background processing in production.

Next Steps

  • See Multi-Agent Design for each agent’s full system prompt, tools, and configuration.
  • See Getting Started to provision the Azure resources and create the Foundry project.