AimmSpike - Semantic Kernel Agent Orchestrator Demo

AimmSpike

Semantic Kernel Agent Orchestrator Demo

~1,400

Lines of C#

5,000+

Python Equivalent

This spike proves that Semantic Kernel can replace DSPy, LangGraph, and LangChain with clean C# code.

How Routing Works

When you send a message, the AgentRouter asks the LLM to classify your query and return a confidence score (0-1).

No ML training required! The LLM naturally understands intent. This replaces 600+ lines of DSPy router code with ~60 lines.

Watch the routing info appear above responses to see this in action.

Available Agents

who_am_i

User profile/identity - populates signature data for emails

chat

General conversation, greetings, questions about capabilities

rag

Document Q&A using vector similarity search (pgvector). Triggers HITL to seed sample docs if empty!

chat_history

Queries about previous conversations in this thread

deal_registration

Create and manage deal registrations (triggers HITL)

email_template

Generate professional email templates

Human-in-the-Loop (HITL)

Some operations require human confirmation. The system will pause and ask for input. Try the deal registration to see this in action!

DataSourceSelection
"Which documents should I search?"

OutputFormatSelection
"How should I format the response?"

ManualConfirmation
"Please confirm this action"

ContextSourceSelection
"Personal or public context?"

PublicResponseAllow
"Can I share this publicly?"

PostResponseFeedback
"Regenerate? Change format?"

Cross-Agent Workflow Demo

This demo shows cross-agent memory - agents can share data within a conversation thread.

Register an Opportunity

First, register a deal or opportunity. The agent will collect info and save it to thread memory.

↓

Provide Contact Info

When prompted, provide a contact name (e.g., "Sarah Johnson").

Wait for the HITL prompt above

↓

Confirm Registration

Click "Yes, register the opportunity" to complete.

Watch for the tip about emails!

↓

Generate Personalized Email

Now ask for an email - it will automatically use the opportunity data!

How it works:

The DealRegistrationAgent saves completed registrations to ThreadMemory. When EmailTemplateAgent runs, it queries memory for recent deals/opportunities and injects that context into the email generation prompt.

Result: Personalized emails with real names and company info!

Architecture

Stack:

• .NET 10 + Semantic Kernel
• PostgreSQL + pgvector
• OpenAI (gpt-4o-mini)
• SSE Streaming

Key Files:

• AgentRouter.cs - LLM routing
• OrchestratorProcess.cs - Workflow
• VectorStore.cs - pgvector RAG
• HilStep.cs - HITL handling
• ThreadMemory.cs - Cross-agent state

Chat Demo

New conversation

Thinking...