IToolMiddleware
Interfaces · AiDotNet.Agentic.Pipeline
A cross-cutting filter around tool/function execution (the AiDotNet analogue of Semantic Kernel’s
function-invocation filter). Middleware can rewrite arguments, require approval, log, time, cache, or
short-circuit a tool call (e.g., deny it) by returning a result without calling next.
Known implementations: ApprovalToolMiddleware, DelegateToolMiddleware
For Beginners
A reusable wrapper around running a tool — ask for confirmation before a risky action, log every call, fix up the inputs, or block it outright — without changing the tool’s code.
How It Works
Applied by wrapping a tool in a MiddlewareAgentTool; the wrapped tool is registered like any
other, so the agent loop runs the middleware transparently whenever the model calls it. This is the hook
for human-in-the-loop tool approval and tool-level guardrails/observability.
Methods
| Method | Summary |
|---|---|
Task<ToolInvocationResult> InvokeAsync(ToolInvocationContext context, ToolPipelineDelegate next, CancellationToken cancellationToken) |
Processes a tool call, optionally calling next to run the tool. |