Semantic Kernel Agents & Workflows (.NET 10)

A compact starter that shows how to:

• Define Semantic Kernel agents with embedded prompts and tool plugins (KernelFunctions).
• Orchestrate typed workflows that chain functions/tools and LLM calls.
• Expose clean entrypoints via dedicated services (agent vs workflow).

What’s in the code

• Program.cs: .NET host composition + logging, registers SK (AddSk), agents/services (AddAgents), then separately invokes an agent service (IIncidentTriageAgentService) and a workflow service (IWorkflowSampleService).
• Infrastructure/Sk/SkConfiguration.cs: builds the Kernel, configures OpenAI chat completion, registers plugins:
  • IncidentTriageTools for the triage agent.
  • WorkflowTools for the workflow (includes delegation to the SummarizationAgent).
• Framework/Agents:
  • AgentBase<TVariables>: handles LLM call, builds ChatHistory, injects system prompt from embedded resource, enables FunctionCalling for authorized KernelFunctions, maps typed input TVariables into KernelArguments.
  • IAgent<TVariables>: only AskAsync(AgentRequest<TVariables>).
  • AgentRequest<TVariables> / AgentResponse: typed request/response contract.
• Framework/Prompting: PromptBuilder reads embedded YAML, deserializes to Prompt, renders via Handlebars + KernelArguments. EmbeddedResourceManager reads embedded resources.
• Framework/Workflow: mini DSL (WorkflowBuilder) + WorkflowRunner + KernelFunctionStep. WorkflowContext.State passes values between steps.
• Sample agents:
  • IncidentTriageAgent (prompt IncidentTriagePrompt.yaml, tools IncidentTriageTools.cs, typed input IncidentTriageInput).
  • SummarizationAgent (prompt SummarizationPrompt.yaml, typed input SummarizationInput).
• Sample services:
  • IncidentTriageAgentService: entrypoint for the triage agent (AgentRequest<IncidentTriageInput> → response).
  • WorkflowSampleService: executes a typed workflow IncidentTriageWorkflow (defined in Samples/Workflow/IncidentTriageWorkflow.cs).

Quick start

1. Prereqs

• .NET 10 SDK.
• OpenAI key: replace open_api_key in Infrastructure/Sk/SkConfiguration.cs (ideally via config/secret).

2. Restore & run

dotnet restore
dotnet run --project PocSKConsole/PocSKConsole.csproj

Agent structure

1) Typed input/output

• Create an input record, e.g., IncidentTriageInput.
• Agent inherits AgentBase<TVariables> with TVariables = input record.
• Call: AskAsync(new AgentRequest<TVariables>("user message", variables), ct);

2) Embedded prompt

• YAML under Samples/MyAgents/<AgentName>/Prompts/<PromptName>.yaml.
• Declared as EmbeddedResource in the csproj.
• Prompt can reference input properties via Handlebars (e.g., {{Incident}}).

3) Tools / KernelFunctions

• Class under Samples/MyAgents/<AgentName>/Tools/....
• Each method annotated [KernelFunction("name")].
• Registered in SK (SkConfiguration) via pluginCollection.AddFromType<YourTools>(pluginName: nameof(YourTools)).
• Agent whitelists functions via AuthorizedKernelFunctions => GetKernelFunctionsForPlugin<YourTools>().

4) Concrete agent

• File e.g., IncidentTriageAgent.cs:
  • PromptFileName → YAML filename (without extension).
  • AuthorizedKernelFunctions → allowed tools.
  • Inherits AgentBase<IncidentTriageInput>.

5) Service entrypoint

• Create a dedicated service, e.g., IncidentTriageAgentService, that takes the agent via DI and exposes a clear method (CallIncidentTriageAgent).
• This service is the “external” entrypoint (API/CLI/worker).

Workflow structure

Typed workflow (existing: IncidentTriageWorkflow)

• Defined in its own class under Samples/Workflow/.
• Builds a WorkflowDefinition via WorkflowBuilder:
  • ThenCallFunction to invoke a KernelFunction (local or LLM-backed).
  • Can store results in WorkflowContext.State.
• WorkflowRunner executes sequentially and returns WorkflowResult.

Tools for workflows

• WorkflowTools exposes operational KernelFunctions (fetch/classify/assign) and delegates to SummarizationAgent for LLM summarization.
• Registered in SK (SkConfiguration).

Service entrypoint

• WorkflowSampleService takes the typed workflow + WorkflowRunner and exposes RunStaticWorkflow.
• This is the “external” entrypoint to orchestrate the workflow.

Add a new agent (step-by-step)

1. Create folder Samples/MyAgents/<NewAgent>/.
2. Create input record Models/<NewInput>.cs.
3. Create prompt YAML in Prompts/<PromptName>.yaml; add as EmbeddedResource in csproj.
4. Create tools in Tools/<NewTools>.cs with [KernelFunction].
5. Create the agent NewAgent.cs inheriting AgentBase<NewInput>, set PromptFileName, whitelist tools.
6. Register in DI (Samples/DI.cs) and in SK (SkConfiguration.cs via AddFromType<NewTools>).
7. Create a dedicated service to expose the agent (<NewAgent>Service) if needed.

Add a new workflow

1. Create a class Samples/Workflow/<NewWorkflow>.cs that builds the WorkflowDefinition via WorkflowBuilder.
2. Add required tools (KernelFunctions) and register them in SkConfiguration.
3. Create a service entrypoint (<NewWorkflow>Service) that takes the typed workflow and WorkflowRunner.
4. Register the workflow and service in DI (Samples/DI.cs).

Tips / gotchas

• Secrets: don’t hardcode the OpenAI key; use config/Secret Manager.
• Prompts: always embed agent prompts and declare them in the csproj.
• Typing: use AgentRequest<TVariables> to force required inputs; the agent turns that into KernelArguments and system prompt automatically.
• Tools: keep tools per agent/workflow separated and registered as distinct plugins.
• Services: keep agent and workflow entrypoints separated to keep responsibilities clear.

What the demo does

• Agent: IncidentTriageAgentService calls the triage agent with an IncidentTriageInput and gets a triage summary (LLM + allowed tools).
• Workflow: WorkflowSampleService runs IncidentTriageWorkflow:
  • fetch_incident (mock) → classify_incident (mock) → summarize_incident (LLM via SummarizationAgent) → assign_incident (mock), chaining state through steps.