Features & Architecture

The ADK provides a robust, composable architecture designed for complex agentic workflows. Build logic efficiently without rewriting boilerplate.

Agent Types

Built-in orchestration patterns determine how a sequence of actions executes:

• LlmAgent: The foundational agent. Handles tool execution, context windows, code execution, and system instructions.
• SequentialAgent: Executes sub-agents sequentially, chaining outputs as inputs.
• ParallelAgent: Runs sub-agents concurrently with identical context, aggregating responses.
• LoopAgent: Runs a sub-agent recursively until a condition is met or an escalation occurs.

Processors Pipeline

LlmAgent uses a highly customizable pipeline to mutate requests and responses before and after LLM dispatch.

Processor	Purpose
InstructionsLlmRequestProcessor	Injects system instructions and few-shot examples.
CodeExecutionRequestProcessor	Enables agents to write and execute code.
RequestConfirmationLlmRequestProcessor	Pauses for human confirmation on sensitive actions.
ContextCacheRequestProcessor	Caches prompts automatically to lower costs.
OutputSchemaRequestProcessor	Forces output to match a strict JSON schema.

Customizing Pipelines

Override the default sequence to inject proprietary logic or logging:

var agent = new LlmAgent(new LlmAgentConfig
{
    Name = "custom_pipeline_agent",
    Model = "gemini-2.5-flash",
    RequestProcessors = [
        BasicLlmRequestProcessor.Instance,
        new MyCustomProcessor(),
        ContentRequestProcessor.Instance
    ]
});

Context Compaction

Prevent context window overflows during long-running flows using automatic compactors.

Compactor	Strategy
Truncation	FIFO - Removes oldest messages.
Token	Evaluates tokens and limits payload size.
Summarization	Uses the LLM to summarize and replace long dialogue blocks.

var agent = new LlmAgent(new LlmAgentConfig
{
    Name = "compact_agent",
    Model = "gemini-2.5-flash",
    ContextCompactors = [new TruncatingContextCompactor(2000)] // Keep last 2000 tokens
});