rskit-agent — turn-based agent loop

rskit-agent runs a bounded agent loop over an injected LLM provider, optional tool registry, and hook registry. It owns turn limits, token budgets, wall-clock budgets, tool-call limits, and context compaction.

Install

[dependencies]
rskit-agent = "0.2.0-alpha.1"
rskit-errors = "0.2.0-alpha.1"
rskit-llm = "0.2.0-alpha.1"
rskit-tool = "0.2.0-alpha.1"

Architecture

graph TD
    ai[rskit-ai]
    llm[rskit-llm]
    tool[rskit-tool]
    hook[rskit-hook]
    errors[rskit-errors]
    agent[rskit-agent]
    loop[turn loop]
    budget[budgets]
    compaction[context policy]
    app[consumer app]

    ai --> agent
    llm --> agent
    tool --> agent
    hook --> agent
    errors --> agent
    agent --> loop
    agent --> budget
    agent --> compaction
    app --> agent

Quick start

use rskit_agent::{Agent, AgentConfig};
use rskit_llm::user;
use rskit_llm_ollama::{self as ollama, Config};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut registry = rskit_llm::Registry::default();
    ollama::register(&mut registry, Config {
        base_url: "http://localhost:11434".into(),
        model: "llama3.2".into(),
        api_key: None,
    })?;
    let provider = registry.build("ollama")?;

    let agent = Agent::new(
        provider,
        AgentConfig {
            system_prompt: "You are concise and operationally precise.".into(),
            model: "llama3.2".into(),
            ..AgentConfig::default()
        },
    );

    let result = agent
        .run(vec![user("Write a two-line release summary.")])
        .await?;

    println!("{}", result.final_message.text());
    Ok(())
}

When to use

Use rskit-agent when you want the bounded turn loop, tool execution, hook points, and context-budget logic without rebuilding orchestration yourself.