Crate lc_agents 

Agent system for building autonomous LLM applications.

Provides core abstractions and implementations for agents.

Core Concepts

• Agent: Responsible for planning, deciding what action to execute next.
• AgentExecutor: Responsible for execution loop (plan -> act -> observe).
• Tool: Callable tools that agents can invoke.

Execution Flow

Input question
    |
Agent.plan() -> AgentAction or AgentFinish
    |
If Action: execute tool -> get observation
    |
Add to intermediate_steps
    |
Loop until AgentFinish returned

Re-exports

pub use executor as base;

pub use orchestrator as orchestration;

pub use adapter::AgentEventRunnable;

pub use adapter::AgentRunnable;

pub use adapter::OrchestratorRunnable;

pub use adaptive_rag::AdaptiveRAG;

pub use adaptive_rag::AdaptiveRAGError;

pub use adaptive_rag::AdaptiveRAGResult;

pub use adaptive_rag::RagDecision;

pub use approval::AllowAll;

pub use approval::ApprovalDecision;

pub use approval::ApprovalHandler;

pub use builder::AgentBuilder;

pub use cache::MemoryCache;

pub use cache::ResponseCache;

pub use crag::CRAGError;

pub use crag::CRAGResult;

pub use crag::CorrectiveRAGAgent;

pub use deep_research::Citation;

pub use deep_research::DeepResearchAgent;

pub use deep_research::ResearchError;

pub use deep_research::ResearchReport;

pub use executor::estimate_step_tokens;

pub use executor::AgentError;

pub use executor::AgentExecutor;

pub use executor::BaseAgent;

pub use executor::BudgetConfig;

pub use executor::BudgetExceeded;

pub use executor::CompactionConfig;

pub use executor::CompactionStrategy;

pub use executor::CompactionTrigger;

pub use function_calling::FunctionCallingAgent;

pub use handoffs::HandoffManager;

pub use hooks::AgentHook;

pub use hooks::ApprovalHook;

pub use hooks::CompletionAction;

pub use hooks::CompletionContext;

pub use hooks::CompletionResult;

pub use hooks::ContentFilterHook;

pub use hooks::ErrorAction;

pub use hooks::HookError;

pub use hooks::LoggingHook;

pub use hooks::PromptInjectionHook;

pub use hooks::StreamAction;

pub use hooks::TokenBudgetHook;

pub use hooks::ToolCallAction;

pub use hooks::ToolCallContext;

pub use hooks::ToolResultContext;

pub use memory_extractor::LlmMemoryExtractor;

pub use orchestrator::parse_review_verdict;

pub use orchestrator::review_envelope;

pub use orchestrator::task_adapter;

pub use orchestrator::FanOutFanIn;

pub use orchestrator::Orchestrator;

pub use orchestrator::ReviewOrchestrator;

pub use orchestrator::ReviewVerdict;

pub use orchestrator::RunContext;

pub use orchestrator::SequentialPipeline;

pub use orchestrator::TaskAdapter;

pub use plan_execute::PlanExecuteAgent;

pub use plan_execute::PlanExecuteError;

pub use policy::ToolPolicy;

pub use policy::ToolRisk;

pub use react::ReActAgent;

pub use resume::FileResumeStore;

pub use resume::MemoryResumeStore;

pub use resume::PendingApproval;

pub use resume::ResumeError;

pub use resume::ResumeStore;

pub use retry::RetryConfig;

pub use streaming::agent_sse_frames;

pub use streaming::encode_sse_frame;

pub use streaming::sse_event_name;

pub use streaming::sse_event_payload;

pub use streaming::AgentEventStream;

pub use streaming::AgentSseRequest;

pub use streaming::AgentStreamEvent;

pub use streaming::SseFrame;

pub use streaming::SseOptions;

pub use streaming::StreamingFunctionCallingAgent;

pub use streaming::ToolCallState;

pub use task::AgentTask;

pub use types::AgentAction;

pub use types::AgentFinish;

pub use types::AgentOutput;

pub use types::AgentStep;

pub use types::ToolInput;

Modules

adapter

AgentRunnable adapter - bridges AgentExecutor to the Runnable trait.

adaptive_rag

Adaptive RAG implementation.

approval

Approval gate (§4.2): asynchronous approval gate before tool execution. Implement ApprovalHandler and inject it via AgentExecutor::with_approval; off by default. Approval gate (§4.2): asynchronous approval gate before tool execution.

builder

AgentBuilder — create an Agent in 3 lines

cache

LLM result cache (P2-1)

crag

Corrective RAG (CRAG) Agent.

deep_research

Deep Research Agent: multi-round research with sub-topic decomposition, parallel search, and comprehensive report synthesis with citations.

executor

Agent base traits and executor implementation.

function_calling

Function calling based agent module.

handoffs

Handoffs multi-agent handoff

hooks

Agent Hook/Middleware system for composable lifecycle interception.

memory_extractor

B4 (v0.22.4): LLM-backed semantic-memory extractor turning a completed turn into durable lc_memory::MemoryItems. B4 (v0.22.4): LLM-backed lc_memory::MemoryExtractor.

metrics

Agent execution metrics (P1-5)

orchestrator

Common trait for high-level orchestrators (P1-1)

plan_execute

Plan-Execute Agent: plan - execute - replan

policy

Tool permission tiering + sandbox gate (P2-9).

react

ReAct Agent implementation

resume

Cross-process resume (§4.2 approval/budget gate): suspend-state persistence + recovery. The framework writes/clears checkpoints (ResumeStore) around approvals; a new process inspects via AgentExecutor::pending_approval and resumes via AgentExecutor::resume. Off by default (no ResumeStore ⇒ no serialization; existing behavior unchanged). Cross-process resume (§4.2 approval/budget gate): suspend-state persistence + recovery.

retry

Retry helpers for LLM calls.

streaming

Streaming Tool Calls

task

Explicit agent task definition (P2-5)

types

Agent related type definitions

Structs

AgentMetrics

Aggregated metrics for a single agent execution.