Planning

Standard LLMs utilize "autoregressive" generation, meaning they predict the next token immediately without explicitly "thinking" ahead. While powerful, this struggles with multi-step reasoning tasks (e.g., "Schedule a meeting, but if X is busy, email Y instead").

The ADK solves this by integrating Planners into the agent execution pipeline.

The ReAct Planner

The most common implementation provided by the ADK is the PlanReActPlanner. ReAct (Reasoning + Acting) fundamentally changes how the model is prompted. It instructs the LLM to output an explicit "Thought" block formulating its internal logic and strategy prior to executing a tool or delivering an action.

using GoogleAdk.Core.Agents;
using GoogleAdk.Core.Planning;
using GoogleAdk.Core.Runner;

var agent = new LlmAgent(new LlmAgentConfig
{
    Name = "planner_agent",
    Model = "gemini-2.5-flash",
    Instruction = "You are a highly logical and meticulous assistant.",

    // Inject the ReAct planner into the middleware pipeline
    Planner = new PlanReActPlanner()
});

var runner = new InMemoryRunner("planning-sample", agent);
var session = await runner.SessionService.CreateSessionAsync(new CreateSessionRequest { AppName = "planning-sample", UserId = "user-1" });

var userMessage = new Content
{
    Role = "user",
    Parts = [new Part { Text = "Plan a weekend itinerary for Paris based on typical weather." }]
};

await foreach (var evt in runner.RunAsync("user-1", session.Id, userMessage))
{
    var text = evt.Content?.Parts?.FirstOrDefault()?.Text;
    if (!string.IsNullOrWhiteSpace(text))
    {
        Console.WriteLine(text);
    }
}

Behind the scenes, the NlPlanningRequestProcessor intercepts the request, appends strict instructional logic requiring "Thought" definitions, and handles the NlPlanningResponseProcessor mapping during response parsing.

Built-in Planner Configuration

For models that natively support specific "thinking" mechanisms directly via their REST API schemas, you can utilize the BuiltInPlanner to pass specific configurations directly to the underlying model adapter.

// Configures the LLM provider's native "thinking" mechanism (e.g., high-reasoning mode)
var builtInPlanner = new BuiltInPlanner(new ThinkingConfig 
{ 
    ThinkingBudget = 1024
});

var agent = new LlmAgent(new LlmAgentConfig
{
    Name = "builtin_planner_agent",
    Model = "gemini-2.5-flash",
    Planner = builtInPlanner
});