Agent Workflow API — Sequential, Parallel

import "github.com/lookatitude/beluga-ai/agent/workflow"

Package workflow provides deterministic workflow agents that orchestrate child agents without LLM reasoning. These agents compose child agents in common execution patterns: sequential pipelines, parallel fan-out, and iterative loops.

Workflow agents implement the [agent.Agent] interface but do not use tools or planners directly. Instead, they delegate to child agents and manage the flow of data between them.

Sequential Agent

SequentialAgent runs child agents in sequence, passing the output of each as the input to the next. This is useful for multi-step pipelines where each stage transforms or enriches the result:

researcher := agent.New("researcher", agent.WithLLM(model))
writer := agent.New("writer", agent.WithLLM(model))
editor := agent.New("editor", agent.WithLLM(model))

pipeline := workflow.NewSequentialAgent("content-pipeline",
    []agent.Agent{researcher, writer, editor},
)

result, err := pipeline.Invoke(ctx, "Write about Go concurrency")

Parallel Agent

ParallelAgent runs child agents concurrently and collects all results. Each child receives the same input. Results are concatenated in order:

analyzer := agent.New("analyzer", agent.WithLLM(model))
reviewer := agent.New("reviewer", agent.WithLLM(model))

parallel := workflow.NewParallelAgent("review-team",
    []agent.Agent{analyzer, reviewer},
)

result, err := parallel.Invoke(ctx, "Review this code")

Loop Agent

LoopAgent runs a single child agent repeatedly until a condition is met or the maximum number of iterations is reached. Configure with WithLoopMaxIterations and WithLoopCondition:

refiner := agent.New("refiner", agent.WithLLM(model))

loop := workflow.NewLoopAgent("iterative-refiner", refiner,
    workflow.WithLoopMaxIterations(3),
    workflow.WithLoopCondition(func(iteration int, result string) bool {
        return strings.Contains(result, "DONE")
    }),
)

result, err := loop.Invoke(ctx, "Refine this draft")

Streaming

All workflow agents support streaming via the Stream method, which yields events from each child agent as they execute. Events include text chunks, tool calls, and completion markers.