swink-agent-patterns

Multi-agent pipeline patterns for swink-agent — compose sequential, parallel, and loop pipelines over an agent factory.

Features

• Pipeline::sequential — chain agents, optionally passing each step's output as the next step's context
• Pipeline::parallel — fan out to N agents concurrently with a MergeStrategy (Concat, First, Fastest { n }, Custom { aggregator })
• Pipeline::loop_ — iterate one agent until an ExitCondition (ToolCalled, OutputContains regex, or MaxIterations)
• PipelineExecutor — drives a pipeline via a pluggable AgentFactory with lifecycle events (PipelineEvent)
• PipelineRegistry — name → pipeline lookup for dynamic dispatch (e.g. routing tools to sub-pipelines)
• PipelineTool — expose a pipeline to the outer agent as a regular AgentTool
• SimpleAgentFactory — register agent builder fns by name for quick prototyping

Quick Start

[dependencies]
swink-agent = "0.8"
swink-agent-patterns = "0.8"
tokio = { version = "1", features = ["full"] }

use std::sync::Arc;
use tokio_util::sync::CancellationToken;
use swink_agent_patterns::{
    Pipeline, PipelineExecutor, PipelineRegistry, SimpleAgentFactory,
};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut factory = SimpleAgentFactory::new();
    factory.register("researcher", || build_research_agent());
    factory.register("writer", || build_writer_agent());

    let registry = Arc::new(PipelineRegistry::new());
    let pipeline = Pipeline::sequential_with_context(
        "research-then-write",
        vec!["researcher".into(), "writer".into()],
    );
    let pipeline_id = pipeline.id().clone();
    registry.register(pipeline);

    let executor = PipelineExecutor::new(Arc::new(factory), registry);
    let output = executor
        .run(&pipeline_id, "Write me a brief on LLM agents.".into(), CancellationToken::new())
        .await?;
    println!("{}", output.final_response);
    Ok(())
}

Architecture

Pipeline is an enum (Sequential, Parallel, Loop) of pure data — each variant references agent steps by name, not by instance. PipelineExecutor materializes agents on demand via the AgentFactory trait, so expensive setup (credentials, tool registration) happens once per run rather than once per step. PipelineEvent emits lifecycle hooks (StepStarted, StepCompleted, PipelineCompleted) so callers can stream progress into a UI or audit log without cracking the executor open.

No unsafe code (#![forbid(unsafe_code)]). Pipelines are Clone + Serialize — you can persist a pipeline definition to disk and recreate it.

Part of the swink-agent workspace — see the main README for workspace overview and setup.