Why should you care?

ERP close workflows need discrete, testable functions instead of one mega-prompt guessing SQL.

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.Plugins/Erp/GlPlugin.cs
public sealed class GlPlugin(IGlRepository gl)
{
    [KernelFunction("fetch_balance")]
    public async Task<decimal> FetchBalanceAsync(string accountCode, int fiscalYear, CancellationToken ct) =>
        await gl.GetBalanceAsync(accountCode, fiscalYear, ct);
}

What happened?

• fetch_balance is the stable tool name models see.
• Returning decimal keeps tool results typed before the LLM narrates them.

Practice next

1. Name functions with stable snake_case IDs for prompts.
2. Keep parameters primitive or DTO-serializable.
3. Handle exceptions inside functions — return error strings, not stack traces.
4. Add optional asOfDate parameter to fetch_balance.
5. Return a BalanceSnapshot record instead of raw decimal.

Remember

KernelFunction methods are AgentNest tools with explicit names. Return small typed results the model can summarize. Test functions independently of chat.

GL balance tool

Finance agent must cite exact ledger balance during variance chat.

Outcome: fetch_balance grounds narratives in SQL, not invented numbers.