AI agent run-loop, prompt builder, and multi-agent orchestration for Garudust.

The centrepiece of this crate is [Agent], which drives the think → tool-call → observe loop until the model signals it is done.

Quick start

use std::sync::Arc;
use garudust_agent::Agent;
use garudust_core::config::AgentConfig;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let config = Arc::new(AgentConfig::default());
    let agent  = Agent::new(config);
    let result = agent.run_once("List files in the current directory").await?;
    println!("{result}");
    Ok(())
}

Architecture

┌──────────────────────────────────────────┐
│                  Agent                   │
│  build_system_prompt()                   │
│  ┌────────────────────────────────────┐  │
│  │  iteration loop                    │  │
│  │  1. call ProviderTransport (LLM)   │  │
│  │  2. dispatch tool calls            │  │
│  │  3. append results to history      │  │
│  │  4. repeat until stop_reason=end   │  │
│  └────────────────────────────────────┘  │
│  persist_session() → SessionDb           │
└──────────────────────────────────────────┘

Skills and self-improvement

When the agent finishes a task it may call write_skill to save a reusable instruction set to ~/.garudust/skills/<name>/SKILL.md. On subsequent runs the skill index is injected into the system prompt so the model can load and apply the skill via skill_view.