Agent Scoping

You can have several agents in one application. By default, every registered component (policy, gate, validator, ...) runs for all agents. With scoping, you can say: "this component runs only for this agent".

Two Ways to Register

// GLOBAL — runs on ALL agents (this is the default)
services.AddAgentQualityGate<JsonQualityGate>();
services.AddAgentResultValidator<MyValidator>();

// SCOPED — runs only on agents that match the condition
services.AddAgentQualityGate<UrlStructureGate>(
    appliesTo: ctx => ctx.AgentName == "UrlResearchAgent");
services.AddAgentInputValidator<PricingFormatValidator>(
    appliesTo: ctx => ctx.AgentName == "PricingAgent");
services.AddAgentResultValidator<DomainValidator>(
    appliesTo: ctx => ctx.AgentName == "DataAgent");

The appliesTo argument is a small function that returns true or false. It receives the IAgentContext, so you can match on any context property — not only the agent name.

Selecting the Agent

The request decides which agent runs, through AgentName. The scoped components compare their condition against this name:

var request = new AgentRequest(
    Goal: "Find pricing URL",
    AgentName: "UrlResearchAgent",      // scoped components match against this
    AllowedToolNames: ["search_web"]);

Input Validation

Input validation is its own step in the pipeline. It checks the input before any work starts:

services.AddAgentInputValidator<ValidUrlInputValidator>(
    appliesTo: ctx => ctx.AgentName == "UrlResearchAgent");

Input validators run after the policies and before the planning. Each validator answers with an InputValidationResult that has two fields: IsValid and Error. If one fails, the run stops immediately. The run then returns an unsuccessful AgentResult that contains the validation error — it does not return the InputValidationResult itself.

Implementation Note

Scoping works with small wrapper classes (FilteredPolicy, FilteredGate, and so on) in Runtime/Filtering. The runtime itself contains no scoping rules.