Agents

#[async_trait]
pub trait Agent: Send + Sync {
    async fn plan(
        &self,
        intermediate_steps: &[(AgentAction, String)],
        inputs: PromptArgs,
    ) -> Result<AgentEvent, AgentError>;

    fn get_tools(&self) -> Vec<Arc<dyn Tool>>;
}

AgentEvent is either Action(Vec<AgentAction>) or Finish(AgentFinish). The executor calls plan() in a loop until Finish or max_iterations.

Uses OpenAI function-calling API. More reliable than ReAct text parsing.

use std::sync::Arc;
use langchainx::{
    agent::{AgentExecutor, OpenAIToolsAgentBuilder},
    chain::Chain,
    llm::openai::OpenAI,
    memory::SimpleMemory,
    prompt_args,
    tools::Tool,
};

let tools: Vec<Arc<dyn Tool>> = vec![Arc::new(MyTool)];

let agent = OpenAIToolsAgentBuilder::new()
    .tools(&tools)
    .llm(OpenAI::default())
    .build()?;

let executor = AgentExecutor::from_agent(agent)
    .with_max_iterations(10)
    .with_memory(SimpleMemory::new().into())
    .with_break_if_error(false);

// AgentExecutor implements Chain — use the Chain API
let result = executor.invoke(prompt_args! {
    "input" => "How many words are in 'hello world'?",
}).await?;
println!("{result}");

Uses text-based ReAct reasoning. Works with any LLM but is less reliable than function calling.

use langchainx::agent::{AgentExecutor, ChatAgentBuilder};

let agent = ChatAgentBuilder::new()
    .tools(&tools)
    .llm(OpenAI::default())
    .build()?;

let executor = AgentExecutor::from_agent(agent)
    .with_max_iterations(10);

AgentExecutor::from_agent(agent)
    .with_max_iterations(10)          // default: Some(10). None = unlimited (dangerous)
    .with_memory(mem.into())          // Arc<Mutex<dyn BaseMemory>>
    .with_break_if_error(true)        // return Err on tool failure instead of continuing

When max_iterations is reached, the executor returns Ok(GenerateResult) with generation = "Max iterations reached" — not an error.

When memory is set, the executor:

1. Injects chat_history into input_variables before each plan() call
2. On Finish, saves user input, tool call messages, and AI response to memory

The required input key is "input" (hardcoded in executor).

let executor = AgentExecutor::from_agent(agent)
    .with_memory(SimpleMemory::new().into());

// First turn
executor.invoke(prompt_args! { "input" => "My name is Alice" }).await?;

// Second turn — executor injects chat_history automatically
executor.invoke(prompt_args! { "input" => "What is my name?" }).await?;