This directory contains specific developer guides for the ADK Python implementation. For the official ADK documentation, visit adk.dev.

• A2aAgentExecutor - Serving an ADK agent behind an A2A server, translating incoming requests into runs and ADK events into task updates.
• A2aRemoteAgentConfig - Controlling the outbound call a RemoteA2aAgent makes to another agent.
• AgentCardBuilder - Deriving the agent card a client reads before sending work, from an agent or workflow you already have.
• to_a2a - Putting an ADK agent on the network as a Starlette application that speaks the Agent2Agent protocol.

• BaseAgent - The foundational base class for custom agents, container orchestrators, and lifecycle callbacks.
• Context - The runtime interface for state, artifacts, memory, credentials, and dynamic execution.
• Creating Agents with Configurations - Building and wiring multi-agent graphs from external YAML configuration files.
• InvocationContext - The runtime dependency container and execution state for a single invocation turn.
• LlmAgent - The primary conversational reasoning agent orchestrating models, tools, and workflows.
• LlmAgent Single-Turn Mode - Guide on using LlmAgent in single-turn mode.
• LlmAgent Task Mode - Guide on using LlmAgent in task mode.
• ManagedAgent - Guide on using ManagedAgent with server-side tools.
• RemoteA2aAgent Task Mode - Guide on using RemoteA2aAgent in task mode.
• RunConfig - Per-invocation streaming, concurrency, session window, telemetry, and live audio settings.

• App - The top-level container binding a root agent to app-wide plugins and configuration.
• EventsCompactionConfig and BaseEventsSummarizer - Keeping a long conversation inside the context window by summarizing older events.
• ResumabilityConfig - Letting an invocation that paused, such as on a long-running tool call, continue from where it stopped.

• BaseArtifactService - Storing binary payloads outside the conversation history, with versioning, user-scoped filenames, and GCS URL generation.

• AuthConfig and authenticated tools - Declaring the credentials a tool needs, and the pause-for-consent handshake.

• ServiceRegistry - Mapping a URI scheme to a factory, so your own session, memory, or artifact store can be selected by URI.
• get_fast_api_app - Serving every agent in a directory over ADK's HTTP API, with room for your own routes, middleware, and lifespan.

• BaseCodeExecutor - Executing model-generated code safely across local, container, GKE, and managed sandbox backends.

• BaseEnvironment and LocalEnvironment - The interface for a place where an agent runs shell commands and keeps files, and the implementation that runs them as local subprocesses.

• ADK exceptions - The exception types ADK raises on its own behalf, and which part of the framework each one comes from.

• AgentEvaluator - Measuring agent quality from inside a pytest suite by replaying recorded conversations and scoring each answer and tool call.
• BaseEvalService and LocalEvalService - Running evaluations that return results as data rather than as a test that passed or failed.
• Efficiency metrics - Reference-free metrics reporting what a run consumed: tool calls, model calls and tokens.
• EvalConfig and the eval config file - The schema of the file that says which metrics score a run and how strict each one is.
• Evaluator - The interface behind the built-in metrics, and how to score a rule that is specific to your agent.

• Event and NodeInfo - Understanding Event and NodeInfo in workflows.
• EventActions - The part of an event that ADK acts on, such as ending a turn or handing the conversation to another agent.
• RequestInput - How to use RequestInput for human-in-the-loop interactions.
• Reserved custom_metadata keys - The Event.custom_metadata keys that only ADK can set, and how restored events are marked.

• Example and ExampleTool - Showing the model worked input and output pairs so that it gets the shape of its own answers right.

• Feature flags - Turning behavior that is not yet stable on or off with the ADK_ENABLE and ADK_DISABLE environment variables.

• Live model callbacks - Inspecting or blocking content on a live bidirectional session.

• AgentEngineEnvironment - Running agent code in an Agent Engine sandbox in your Google Cloud project, which pauses between uses instead of being deleted.
• BigQueryToolset - Exploring and querying BigQuery, and the write_mode setting that decides what the agent may change.
• CrewaiTool - Wrapping a CrewAI tool so an ADK agent can call it.
• DaytonaEnvironment - Running agent code in a Daytona hosted sandbox instead of on your machine.
• E2BEnvironment - Running agent code in an E2B hosted sandbox, and what happens when the sandbox expires.
• FirestoreSessionService - A durable multi-process session store built on Firestore documents and transactions.
• GCSToolset and GCSAdminToolset - Giving an agent access to Cloud Storage objects and buckets, read-only until you say otherwise.
• LangchainTool - Wrapping a LangChain tool so an ADK agent can call it.
• Model Armor - Screening user input and model output with Google Cloud Model Armor.
• MongoDbMemoryService - Storing session events in a MongoDB collection and recalling them by keyword match.
• MongoDbSessionService - Sharing sessions across processes through MongoDB, with transactional writes where the deployment supports them.
• MongoDbToolset - Vector and hybrid search over a MongoDB database, with the query text embedded on the way through.
• RedisSessionService - Sharing sessions across processes through Redis, including the expiry every other backend lacks.
• TypesafeClassifier - Answering Choice, Binary, and Score questions with TypeSafe's Jev decision models instead of a model turn.

• AntigravityAgent - Runs a Google Antigravity SDK agent as an ADK agent node.
• ClassifierNode - Routing a workflow on a classifier's answer to one Choice or Binary question.
• LlmClassifier - Answering Choice, Binary, and Score questions with any model, in one structured-output call.

• Live tools - Asynchronous background execution and response scheduling for Gemini Live agents.
• CascadeLive - A live-composing model wrapping a text reasoner in speech-to-text and text-to-speech transforms.
• GeminiSTT - Recognizing a CascadeLive agent's user speech with Gemini streaming transcription, and when the user's turn ends.
• GeminiTTS - Speaking a CascadeLive agent's replies with Gemini text-to-speech, sentence by sentence and without gaps.
• LiveRequestQueue - Streaming content, realtime audio, and stream control signals to live agents.

• BaseMemoryService - Storing finished sessions and recalling them from later conversations.

• BaseLlm and LLMRegistry - The model interface, how a model name resolves to an implementation, and how to plug in your own.
• FallbackModel - Wrapping an ordered list of models and moving to the next one when a call fails.
• ServiceTier - Choosing serving tiers for Interactions API calls, including deferred execution on off-peak capacity.

• AgentOptimizer and Sampler - Rewriting an agent's instruction automatically, scoring candidate prompts against an evaluation set and keeping the better one.

• BasePlanner - Guiding model execution with structured planning instructions, thinking configurations, and Plan-Re-Act thought tagging.

• ForceEscalationPlugin - Enforcing deterministic first-turn, repeated-tool-call, and pre-tool gating escalations to ModelConsultTool.
• ModelConsultObservabilityPlugin - Measuring ModelConsultTool escalations, advisor and executor tokens, latency, and estimated cost for every invocation.
• ReflectAndRetryModelPlugin - Self-healing, concurrent-safe error recovery for model failures.
• ReflectAndRetryToolPlugin - Self-healing, concurrent-safe error recovery for tool failures.
• ToolCallIntegrityPlugin - Optional HMAC integrity check for stored function call arguments.

• Runner and InMemoryRunner - Managing session lifecycles, state resolution, and streaming agent execution events.
• Runner Execution Cancellation - Halting agent and workflow execution cleanly using abort signals.
• Runner Live Streaming - Real-time bidirectional audio/text streaming and non-blocking background tool execution with Gemini Multimodal Live API.
• Runner.rewind_async - Undoing turns of a conversation and restoring state and artifacts to an earlier point.

• Credentials Encryption - Securely encrypting sensitive session credentials using Google Cloud KMS.

• Session and BaseSessionService - The session lifecycle, state scoping, and choosing a session service.
• State - Session state and the app:, user:, and temp: prefixes that decide what is shared and what is stored.

• Skill, Frontmatter, and Resources - The SKILL.md file format, and the folder of instructions, reference documents, and scripts an agent pulls in only when it is relevant.
• SkillRegistry - The interface behind a searchable catalog of skills that an agent discovers at runtime.

• TelemetryConfig - What ADK puts in its OpenTelemetry traces, and whether the text of prompts and replies is copied onto exported spans.

• FunctionTool - Wrapping Python functions and generators as agent tools with argument validation, progress streaming, and confirmation.
• Using Context in a tool - Reaching session state, artifacts, memory, and credentials from inside a tool.
• BaseTool - Subclassing the tool base class to write a declaration yourself or to change the model request.
• BaseToolset - Giving an agent a group of tools that ADK asks for on every run, and filtering which ones the model sees.
• LongRunningFunctionTool - Tools whose result arrives later, such as an approval or a batch job, and how to send that result back.
• ToolConfirmation - Pausing a tool call until a person approves it, and passing their answer back to the agent.
• AgentTool - Calling another agent as a tool in a separate session, and how that compares with a single-turn sub-agent.
• McpToolset - Connecting an agent to a Model Context Protocol server and using the tools it publishes.
• SkillToolset - Giving an agent skills that it loads only when a request calls for them.
• ModelConsultTool and ModelConsultContextConfig - Escalating hard decisions mid-generation to a stronger advisor model, with per-turn and session budgets.
• Node as tool - Exposing workflows and deterministic nodes as agent tools with isolated runtime branching and resume support.
• to_mcp_server - Expose an ADK agent as an MCP server so any MCP host can drive it as a single tool (the MCP counterpart of to_a2a).
• ToolArgsMatcher - Comparing a tool call's arguments with expected ones per argument, by equality, case, numeric tolerance, substring, or a custom matcher.

• inject_session_state - Substituting session state values and artifact contents into an instruction string.

• BaseNode - The foundational base class and configuration settings for all workflow nodes.
• Node and @node - Wrapping functions, agents, and tools as workflow steps, overriding node settings per graph, and subclassing Node.
• Workflow - Graph-based orchestration of complex, multi-step agent interactions.
• Workflow Graphs - Understanding nodes, edges, and graph structures in workflows.
• Function Nodes - Wrapping Python functions and generators as workflow nodes.
• JoinNode - Synchronizing parallel execution paths in workflows.
• RetryConfig - Configuring retry policies for resilient workflow nodes.
• ParallelWorker - Processing lists of items concurrently in workflows.
• Dynamic Nodes - Scheduling and executing nodes dynamically at runtime.