Track
Reasoning Agents (Azure AI Foundry)
Project Name
CertPrepAgents — Multi-Agent Certification Coach
GitHub Username
@ANcpLua
Repository URL
https://github.com/ANcpLua/agentsleague-starter-kits/tree/main/track-2-reasoning-agents
Project Description
CertPrepAgents is a five-agent system built with the Microsoft Agent Framework (.NET 10) that coaches students through Microsoft certification exam preparation using genuine multi-step reasoning.
The pipeline: A Curator Agent searches the live Microsoft Learn API for official training modules. A Study Plan Generator structures those modules by exam domain weight. An Assessment Agent — powered by the o4-mini reasoning model — generates scenario-based exam questions and grades answers with educational reasoning. A Certification Planner analyzes per-domain scores and decides pass or remediate. An Engagement Agent sends the student actionable study communications via email.
The key differentiator is the remediation loop. When a student fails, the system doesn't restart — it diagnoses which specific domains are weak, references the exact questions answered incorrectly, estimates remediation hours scaled to the gap size, and loops back through the pipeline targeting only those domains. Up to three adaptive attempts before final engagement.
Technical highlights: Content safety middleware blocks prompt injection before any agent processes input. OpenTelemetry distributed tracing makes every agent decision auditable as spans. All agents communicate via typed JSON contracts with source-generated serialization. The entire system runs on GitHub Models free tier (gpt-4o-mini + o4-mini) — zero Azure spend to reproduce.
This implements four reasoning patterns: Planner-Executor (Study Plan plans, Assessment executes), Critic/Verifier (Planner verifies Assessment output), Self-reflection (remediation loop), and Role-based specialization (5 agents with clean contracts). Built with GitHub Copilot in Agent, Edit, and Plan modes.
Demo Video or Screenshots
Screenshots will be added shortly after submission
Primary Programming Language
C#/.NET
Key Technologies Used
- Microsoft Agent Framework (Microsoft.Agents.AI v1.0.0-preview) — agent orchestration with sequential workflows
- GitHub Models (free tier) — gpt-4o-mini for standard agents, o4-mini for reasoning
- Microsoft Learn API — live content retrieval (learn.microsoft.com/api/search)
- OpenTelemetry — distributed tracing with OTLP export
- .NET 10 / C# 14
- GitHub Copilot — used throughout development (Agent, Edit, Plan modes)
Submission Type
Individual
Team Members
No response
Submission Requirements
Quick Setup Summary
- Clone:
git clone https://github.com/ANcpLua/agentsleague-starter-kits.git
- Navigate:
cd track-2-reasoning-agents/src/CertPrepAgents
- Set token:
dotnet user-secrets set "GITHUB_TOKEN" "your-github-pat"
- Build and run:
dotnet run
- Open DevUI:
http://localhost:5180
Optional OTel: Run an OTLP collector on localhost:4317 (e.g., Aspire Dashboard)
Technical Highlights
- 5-agent sequential workflow with adaptive remediation loop (up to 3 attempts targeting only weak domains)
- o4-mini reasoning model for higher-quality assessment generation and grading
- Live Microsoft Learn API integration — real URLs from learn.microsoft.com/api/search, not hallucinated
- Content safety middleware blocking prompt injection at the HTTP layer before any agent processes input
- Typed JSON contracts between agents with source-generated serialization (CuratorOutput, StudyPlan, AssessmentResult, PlannerDecision, RemediationPlan)
- OpenTelemetry tracing for full reasoning chain auditability (Curator → Study Plan → Assessment → Planner → Engagement)
- Zero Azure spend — runs entirely on GitHub Models free tier (gpt-4o-mini + o4-mini)
- Clean architecture: Contracts/, Prompts/, Tools/, Middleware/ separation
Challenges & Learnings
- Microsoft Agent Framework is in early preview — learned to work with AddWorkflow and AgentWorkflowBuilder.BuildSequential for multi-agent orchestration, and the middleware pattern for implementing the remediation loop via .AsBuilder().Use()
- Reasoning model (o4-mini) required different prompting strategy than standard models — structured JSON output with educational reasoning for each graded answer
- The remediation loop needed careful design to avoid restarting from scratch — enriching context with weak-domain information from the Planner's analysis was the key insight that made targeted remediation work
- Content safety at the middleware level (before agents) vs per-agent guardrails proved more robust and eliminated redundant safety prompts across all 5 agents
Contact Information
anfh22@outlook.com — Alexander Nachtmann
Country/Region
Austria
Track
Reasoning Agents (Azure AI Foundry)
Project Name
CertPrepAgents — Multi-Agent Certification Coach
GitHub Username
@ANcpLua
Repository URL
https://github.com/ANcpLua/agentsleague-starter-kits/tree/main/track-2-reasoning-agents
Project Description
CertPrepAgents is a five-agent system built with the Microsoft Agent Framework (.NET 10) that coaches students through Microsoft certification exam preparation using genuine multi-step reasoning.
The pipeline: A Curator Agent searches the live Microsoft Learn API for official training modules. A Study Plan Generator structures those modules by exam domain weight. An Assessment Agent — powered by the o4-mini reasoning model — generates scenario-based exam questions and grades answers with educational reasoning. A Certification Planner analyzes per-domain scores and decides pass or remediate. An Engagement Agent sends the student actionable study communications via email.
The key differentiator is the remediation loop. When a student fails, the system doesn't restart — it diagnoses which specific domains are weak, references the exact questions answered incorrectly, estimates remediation hours scaled to the gap size, and loops back through the pipeline targeting only those domains. Up to three adaptive attempts before final engagement.
Technical highlights: Content safety middleware blocks prompt injection before any agent processes input. OpenTelemetry distributed tracing makes every agent decision auditable as spans. All agents communicate via typed JSON contracts with source-generated serialization. The entire system runs on GitHub Models free tier (gpt-4o-mini + o4-mini) — zero Azure spend to reproduce.
This implements four reasoning patterns: Planner-Executor (Study Plan plans, Assessment executes), Critic/Verifier (Planner verifies Assessment output), Self-reflection (remediation loop), and Role-based specialization (5 agents with clean contracts). Built with GitHub Copilot in Agent, Edit, and Plan modes.
Demo Video or Screenshots
Screenshots will be added shortly after submission
Primary Programming Language
C#/.NET
Key Technologies Used
Submission Type
Individual
Team Members
No response
Submission Requirements
Quick Setup Summary
git clone https://github.com/ANcpLua/agentsleague-starter-kits.gitcd track-2-reasoning-agents/src/CertPrepAgentsdotnet user-secrets set "GITHUB_TOKEN" "your-github-pat"dotnet runhttp://localhost:5180Optional OTel: Run an OTLP collector on localhost:4317 (e.g., Aspire Dashboard)
Technical Highlights
Challenges & Learnings
Contact Information
anfh22@outlook.com — Alexander Nachtmann
Country/Region
Austria