.NET: [Bug]: Agent Skill throws "System.InvalidOperationException"

Description

I'm exploring the examples of the Microsoft Agent Framework and have a question or maybe found a problem or issue. ;)

When I execute the example Agent_Step03_ClassBasedSkills using my local AI (Model qwen3.8-27b) with the OpenAI API, the Agent returns something like this:

Converting units with class-based skills
------------------------------------------------------------
Agent:

The conversion script isn't working, so I calculated these using the conversion table's formula (result = value × factor):

## Marathon: 26.2 miles ␦ kilometers
- **26.2 × 1.60934 = 42.16 km** (42.164708, rounded)

So a marathon is about **42.16 kilometers** - commonly rounded to 42.2 km.

## 75 kilograms ␦ pounds
- **75 × 2.20462 = 165.35 lbs** (165.3465, rounded)

So 75 kilograms is about **165.35 pounds**.

The important part is "The conversion script isn't working, so I calculated these using the conversion table's formula (result = value × factor)"

In the output windows I see a couple of exceptions like this:

Exception thrown: 'System.InvalidOperationException' in Microsoft.Agents.AI.dll
Inline skill scripts expect arguments as a JSON object but received a JSON element of kind 'String'.

Debugging further gives me the AgentInlineSkillScript and there I found

 private static AIFunctionArguments ConvertToFunctionArguments(JsonElement? arguments)
    {
        if (arguments is null ||
            arguments.Value.ValueKind == JsonValueKind.Null ||
            arguments.Value.ValueKind == JsonValueKind.Undefined)
        {
            return [];
        }

        if (arguments.Value.ValueKind != JsonValueKind.Object)
        {
            throw new InvalidOperationException(
                $"Inline skill scripts expect arguments as a JSON object but received a JSON element of kind '{arguments.Value.ValueKind}'.");
        }

        var dict = new Dictionary<string, object?>();
        foreach (var property in arguments.Value.EnumerateObject())
        {
            dict[property.Name] = property.Value;
        }

        return new AIFunctionArguments(dict);
    }

So, no matter what the LLM is using as parameter calling the skill script, as long as this is no JSON-Object, this function throws an exception.

So my question is:
how can I call the skill with the two double values as expected, so that the LLM can use the skill and does not run into a failure?

Code Sample

/// <summary>
    /// Converts a value by the given factor.
    /// </summary>
    [AgentSkillScript("convert")]
    [Description("Multiplies a value by a conversion factor and returns the result as JSON.")]
    private static string ConvertUnits(double value, double factor)
    {
        double result = Math.Round(value * factor, 4);
        return JsonSerializer.Serialize(new { value, factor, result });
    }

Error Messages / Stack Traces

Package Versions

Microsoft.Agents.AI: 1.20.0; Microsoft.Agents.AI.OpenAI: 1.20.0;

.NET Version

.NET 10

Additional Context

No response

s-kip

👋 Hi s-kip — thanks for the report!

Your exception explicitly says the model supplied arguments as a JSON String, while the loaded skill schema requires an object. The correct run_skill_script payload is {"skillName":"unit-converter","scriptName":"convert","arguments":{"value":26.2,"factor":1.60934}}; with that shape, your two values produce 42.1647. Your local model appears to be quoting/stringifying the arguments object instead of emitting it directly.

If you still believe there is a framework issue here, please reply with the exact import/class or method you are using, the version, and a minimal snippet that shows the failure. 🙏

I debugged the ConvertToFunctionArguments of the AgentInlineSkillScript.cs and this is an exmaple of the parameter arguments that was used:
"{"value": 26.2, "factor": 1.60934}"

So is the LLM the problem, because it uses the wrong values for the parameter arguments?
If yes, what can I do to "tell" my LLM what to use instead?

Your debugging is correct — the model is the immediate cause. Your local model is emitting arguments double-encoded: a JSON string containing {"value": 26.2, "factor": 1.60934} rather than the object itself. Smaller/local models do this fairly often; GPT-class models usually don't.

The run_skill_script tool declares arguments as JsonElement? (AgentSkillsProvider.cs), which gives an unconstrained schema, so nothing at the tool-call level tells the model it must be an object — and the default marshaler then rejects anything that isn't one.

You don't have to patch the framework though: every skill type accepts an argumentMarshaler, including AgentClassSkill<TSelf>. Adding one to the sample skill makes it tolerate the stringified form:

internal sealed class UnitConverterSkill : AgentClassSkill<UnitConverterSkill>
{
    public UnitConverterSkill() : base(LenientArguments) { }

    private static AIFunctionArguments LenientArguments(JsonElement? arguments)
    {
        if (arguments is not { } el || el.ValueKind is JsonValueKind.Null or JsonValueKind.Undefined)
        {
            return [];
        }

        // Some models stringify the arguments object; unwrap it once.
        if (el.ValueKind == JsonValueKind.String)
        {
            el = JsonSerializer.Deserialize<JsonElement>(el.GetString()!);
        }

        if (el.ValueKind != JsonValueKind.Object)
        {
            throw new InvalidOperationException($"Expected a JSON object, received '{el.ValueKind}'.");
        }

        var dict = new Dictionary<string, object?>();
        foreach (var p in el.EnumerateObject())
        {
            dict[p.Name] = p.Value;
        }

        return new AIFunctionArguments(dict);
    }
}

If you'd also like to nudge the model in the prompt, AgentSkillsProviderOptions.SkillsInstructionPrompt replaces the default skills instruction block — you can append something like "pass arguments to run_skill_script as a JSON object, never as a JSON-encoded string". Note that this needs the collection constructor rather than the params one used in the sample:

var skillsProvider = new AgentSkillsProvider(
    [unitConverter],
    new AgentSkillsProviderOptions { SkillsInstructionPrompt = /* customized prompt */ });

Expect that to help but not be fully reliable on a small local model; the marshaler is the deterministic fix.

For the maintainers: it may be worth having the default marshaler unwrap a JSON string that contains an object (a handful of lines), since double-encoded tool arguments are common with OpenAI-compatible local models, and mentioning the argumentMarshaler escape hatch in the exception message — right now nothing points users to it.

Atharva Vichare (@atty57)

Thx Atharva Vichare (@atty57) for clarifying on this.
I tried to set the SkillInstructionPrompt, without success.

But it worked with the custom marshaler you wrote 👍

In my opinion, it would be great if the default marshaler handles this automatically.
Or the exception should contain a hint to the solution with the custom marshaler.
Because,as you already mentioned, nothing points me to this solution.

s-kip

s-kip, thanks for reporting this issue. Agent Framework should not automatically “heal” malformed JSON, as guessing the intended structure can be ambiguous, hide model/provider interoperability problems, and produce unexpected behavior. Providers should emit arguments that conform to the declared tool schema, while applications can use a custom argument marshaler when model-specific normalization is required.

We’ve addressed your suggestion by updating the exception message to point users to the custom argument marshaler.

Navigation Menu

Repository navigation

.NET: [Bug]: Agent Skill throws "System.InvalidOperationException" #7989

Description

Code Sample

Error Messages / Stack Traces

Package Versions

.NET Version

Additional Context

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

.NET: [Bug]: Agent Skill throws "System.InvalidOperationException" #7989

Description

Description

Code Sample

Error Messages / Stack Traces

Package Versions

.NET Version

Additional Context

Activity

github-actions commented on Sep 1, 2026

s-kip commented on Sep 1, 2026

atty57 commented on Sep 1, 2026

s-kip commented on Sep 2, 2026

SergeyMenshykh commented on Sep 7, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions