Golem Add Agent Rust

Adding a New Agent to a Rust Golem Component

Overview

An agent is a durable, stateful unit of computation in Golem. Each agent type is defined as a trait annotated with #[agent_definition] and implemented on a struct annotated with #[agent_implementation].

Steps

1. Create the agent module — add a new file src/<agent_name>.rs
2. Define the agent trait — annotate with #[agent_definition]
3. Implement the agent — annotate with #[agent_implementation]
4. Re-export from lib.rs — add mod <agent_name>; and pub use <agent_name>::*;
5. Build — run golem build to verify

Agent Definition

use golem_rust::{agent_definition, agent_implementation};

#[agent_definition]
pub trait MyAgent {
    // Constructor parameters form the agent's identity.
    // Two agents with the same parameters are the same agent.
    fn new(name: String) -> Self;

    // Agent methods — can be sync or async
    fn get_count(&self) -> u32;
    fn increment(&mut self) -> u32;
    async fn fetch_data(&self, url: String) -> String;
}

struct MyAgentImpl {
    name: String,
    count: u32,
}

#[agent_implementation]
impl MyAgent for MyAgentImpl {
    fn new(name: String) -> Self {
        Self { name, count: 0 }
    }

    fn get_count(&self) -> u32 {
        self.count
    }

    fn increment(&mut self) -> u32 {
        self.count += 1;
        self.count
    }

    async fn fetch_data(&self, url: String) -> String {
        // Use wstd::http for HTTP requests
        todo!()
    }
}

Custom Types

All parameter and return types must implement the Schema trait. For custom types, derive it along with IntoValue and FromValueAndType:

use golem_rust::Schema;
use serde::{Serialize, Deserialize};

#[derive(Clone, Schema, Serialize, Deserialize)]
pub struct MyData {
    pub field1: String,
    pub field2: u32,
}

Returning Failures

Agent methods should distinguish between domain errors (expected failure outcomes) and uncaught errors:

• Uncaught errors (panics, unwrapped Err, etc.) are not returned to the caller as a failed invocation. Golem treats them as crashes: the invocation is retried according to the agent's retry policy, and if the retries are exhausted the agent itself becomes failed.
• Domain errors that the caller should observe as a normal failure result must be expressed in the method's return type using Result<T, E>. The error type E must derive Schema like any other custom type.

use golem_rust::{agent_definition, agent_implementation, Schema};
use serde::{Serialize, Deserialize};

#[derive(Clone, Debug, Schema, Serialize, Deserialize)]
pub enum WithdrawError {
    InsufficientFunds { available: u64 },
    AccountClosed,
}

#[agent_definition]
pub trait Wallet {
    fn new(owner: String) -> Self;

    fn withdraw(&mut self, amount: u64) -> Result<u64, WithdrawError>;
}

struct WalletImpl {
    owner: String,
    balance: u64,
    closed: bool,
}

#[agent_implementation]
impl Wallet for WalletImpl {
    fn new(owner: String) -> Self {
        Self { owner, balance: 0, closed: false }
    }

    fn withdraw(&mut self, amount: u64) -> Result<u64, WithdrawError> {
        if self.closed {
            return Err(WithdrawError::AccountClosed);
        }
        if amount > self.balance {
            return Err(WithdrawError::InsufficientFunds { available: self.balance });
        }
        self.balance -= amount;
        Ok(self.balance)
    }
}

Returning Err(WithdrawError::...) completes the invocation successfully — the caller receives the error as a value. Panicking with unwrap(), expect(), panic!(), or returning from ? on an unrelated error type will instead trigger a retry and eventually fail the whole agent.

Key Constraints

• All agent method parameters are passed by value (no references)
• All custom types need Schema derive (plus IntoValue and FromValueAndType, which Schema implies)
• Constructor parameters form the agent identity — two agents with the same parameters are the same agent
• Agents are created implicitly on first invocation — no separate creation step
• Invocations are processed sequentially in a single thread — no concurrency within a single agent
• Never use block_on — all agent methods run in an async context. Use async methods and .await instead of blocking on futures