Why should you care?

Regulated customers require traceable rationale; hidden jumps from question to action fail hospital and finance reviews.

See it live — copy this example

Paste into an ASP.NET Core 8+ / AgentNest project, then run with dotnet run (set your API keys in user-secrets).

// AgentNest.Agents/Reasoning/VarianceReasoner.cs
public sealed class VarianceReasoner(IChatClient chat)
{
    public async Task<VarianceAnalysis> AnalyzeAsync(string glLine, decimal delta, CancellationToken ct)
    {
        var prompt = $"""
            GL line: {glLine}, variance: {delta:C}.
            Step 1: list plausible causes.
            Step 2: rank by likelihood.
            Step 3: recommend one ERP action.
            Return JSON with steps[] and recommendation.
            """;
        var json = (await chat.GetResponseAsync([prompt], cancellationToken: ct)).Text;
        return JsonSerializer.Deserialize<VarianceAnalysis>(json)!;
    }
}

What happened?

• The prompt forces numbered reasoning before a recommendation.
• VarianceAnalysis captures steps for UI display and auditing, not just final text.

Practice next

1. Define VarianceAnalysis with a Steps list and Recommendation property.
2. Configure System.Text.Json camelCase naming for model JSON.
3. Validate deserialized output; return 422 if JSON is malformed.
4. Add a ConfidenceScore field and reject recommendations below 0.6.
5. Use ChatResponseFormat.Json for stricter model output.

Remember

Reasoning prompts ask the model to show intermediate steps. Structured DTOs make ERP and healthcare rationales reviewable. Always validate JSON from models before persisting or acting.

ERP variance review

Controllers reject black-box AI summaries during SOX audits.

Outcome: VarianceReasoner returns step arrays auditors can replay against ledger data.