Crate adk_agent 

adk-agent

Agent implementations for ADK (LLM, Custom, Workflow agents).

Overview

This crate provides ready-to-use agent implementations:

• LlmAgent - Core agent powered by LLM reasoning
• CustomAgent - Define custom logic without LLM
• SequentialAgent - Execute agents in sequence
• ParallelAgent - Execute agents concurrently, with optional SharedState coordination
• LoopAgent - Iterate until exit condition
• ConditionalAgent - Branch based on conditions
• TeamSpec - Validate and compile a portable team from existing agents

What’s New in 0.6.0

• ParallelAgent::with_shared_state(): Opt-in builder method that creates a fresh SharedState per run() invocation, enabling sub-agents to exchange data via set_shared/get_shared/wait_for_key through the context chain.
• AgentToolContext delegation: Tools executed by LlmAgent now have access to shared_state() via the context chain, enabling tool-level coordination in parallel workflows.
• Tool confirmation: LlmAgentBuilder::require_tool_confirmation() and require_tool_confirmation_for_all() for human-in-the-loop tool authorization.

Quick Start

use adk_agent::LlmAgentBuilder;
use std::sync::Arc;

// LLM Agent requires a model (from adk-model)
// let agent = LlmAgentBuilder::new("assistant")
//     .description("Helpful AI assistant")
//     .model(Arc::new(model))
//     .build()?;

Workflow Agents

Combine agents for complex workflows:

// Sequential: A -> B -> C
let seq = SequentialAgent::new("pipeline", vec![a, b, c]);

// Parallel: A, B, C simultaneously
let par = ParallelAgent::new("team", vec![a, b, c]);

// Loop: repeat until exit
let loop_agent = LoopAgent::new("iterator", vec![worker]).with_max_iterations(10);

Guardrails (optional)

Enable the guardrails feature for input/output validation:

use adk_agent::{LlmAgentBuilder, guardrails::{GuardrailSet, ContentFilter, PiiRedactor}};

let input_guardrails = GuardrailSet::new()
    .with(ContentFilter::harmful_content())
    .with(PiiRedactor::new());

let agent = LlmAgentBuilder::new("assistant")
    .input_guardrails(input_guardrails)
    .build()?;

Re-exports

pub use compaction::LlmEventSummarizer;

pub use guardrails::GuardrailSet;

pub use team::BlackboardHistoryPolicy;

pub use team::BlackboardPolicy;

pub use team::BlackboardSchedule;

pub use team::BlackboardSpec;

pub use team::BlackboardTransition;

pub use team::CircuitBreakerPolicy;

pub use team::CompiledBlackboardTeam;

pub use team::CompiledTeam;

pub use team::RelationshipApprovalPolicy;

pub use team::RelationshipFailureStrategy;

pub use team::RelationshipKind;

pub use team::RelationshipPolicy;

pub use team::ResolvedTeamMember;

pub use team::StaticTeamAgentRegistry;

pub use team::TEAM_EDGE_ID_KEY;

pub use team::TEAM_EXECUTION_STATE_KEY;

pub use team::TEAM_ROOT_INVOCATION_KEY;

pub use team::TeamAgentDescriptor;

pub use team::TeamAgentHealth;

pub use team::TeamAgentRegistry;

pub use team::TeamArchitectureTemplate;

pub use team::TeamBudget;

pub use team::TeamContextPolicy;

pub use team::TeamEdgeExecution;

pub use team::TeamError;

pub use team::TeamExecutionAnalysis;

pub use team::TeamExecutionSnapshot;

pub use team::TeamExecutionStatus;

pub use team::TeamExecutionUsage;

pub use team::TeamFailurePolicy;

pub use team::TeamHistoryPolicy;

pub use team::TeamLifecycleContext;

pub use team::TeamLifecycleDecision;

pub use team::TeamLifecycleHook;

pub use team::TeamLifecycleOutcome;

pub use team::TeamLifecyclePhase;

pub use team::TeamManagerBranch;

pub use team::TeamMemberSpec;

pub use team::TeamPolicy;

pub use team::TeamRegistryRequirement;

pub use team::TeamRelationship;

pub use team::TeamReplayError;

pub use team::TeamResumePlan;

pub use team::TeamResumePolicy;

pub use team::TeamRuntimeError;

pub use team::TeamSpec;

pub use team::TeamStateMergePolicy;

pub use team::TeamTerminationPolicy;

pub use team::WorkflowArchitectureTemplate;

pub use team::analyze_team_execution;

pub use team::validate_team_replay;

pub use tool_call_markup::normalize_content;

pub use tool_call_markup::normalize_option_content;

Modules

compaction

LLM-based event summarizer for context compaction.

guardrails

Guardrail integration for LlmAgent

team

Portable, validated multi-agent team composition.

tool_call_markup

XML-based tool call markup parsing.

Structs

ConditionalAgent

Rule-based conditional routing agent.

CustomAgent

An agent with a user-defined async handler function.

CustomAgentBuilder

Builder for constructing a CustomAgent.

LlmAgent

An LLM-powered agent that orchestrates tool calls and sub-agent delegation.

LlmAgentBuilder

Builder for constructing an LlmAgent with all configuration options.

LlmConditionalAgent

LLM-based intelligent conditional routing agent.

LlmConditionalAgentBuilder

Builder for constructing an LlmConditionalAgent.

LoopAgent

Loop agent executes sub-agents repeatedly for N iterations or until escalation

ParallelAgent

Parallel agent executes sub-agents concurrently

SequentialAgent

Sequential agent executes sub-agents once in order

Constants

DEFAULT_LOOP_MAX_ITERATIONS

Default maximum iterations for LoopAgent when none is specified. Prevents infinite loops from consuming unbounded resources.

DEFAULT_MAX_ITERATIONS

Default maximum number of LLM round-trips (iterations) before the agent stops.

DEFAULT_TOOL_TIMEOUT

Default tool execution timeout (5 minutes).

Traits

Agent

The fundamental trait for all ADK agents.

Functions

extract_typed

Extract a typed value from agent events.

Type Aliases

AfterToolCallbackFull

Rich after-tool callback that receives the tool, arguments, and response.

OnToolErrorCallback

Callback invoked when a tool execution fails (after retries are exhausted).