Workflow Agents

Workflow agents orchestrate multiple agents to solve complex problems that require coordination, parallel processing, or iterative refinement. Think of them as the conductors of your agent orchestra - they don't generate content themselves, but control how and when other agents execute.

Unlike LLM Agents that use language models for reasoning, workflow agents operate with deterministic logic to ensure predictable, reliable execution patterns. This makes them perfect for building robust multi-step processes where you need guaranteed execution order and consistent results.

Workflow vs LLM Agents

Workflow Agents manage execution flow with predictable patterns (sequential → parallel → loop → LangGraph).

LLM Agents use language models for dynamic reasoning and content generation.

Best together: Use workflow agents to orchestrate LLM agents for complex, multi-step processes.

Workflow Types

Sequential Agent

Execute agents one after another in a specific order

Parallel Agent

Run multiple agents simultaneously for faster processing

Loop Agent

Repeat agent execution until conditions are met

LangGraph Agent

Complex workflows with conditional branching and dynamic routing

Key Benefits

Workflow agents solve real problems that single agents can't handle effectively:

🔄 Reliable Process Control
Guarantee execution order and patterns without relying on LLM decision-making. Perfect for production workflows where consistency matters.

⚡ Performance Optimization
Run independent tasks in parallel or iterate until quality standards are met. Scale processing power where it's needed most.

🧩 Composable Complexity
Break complex problems into manageable steps. Each sub-agent can focus on what it does best while the workflow manages coordination.

🎯 Specialized Orchestration
Combine different agent types - use LLM agents for reasoning, tool agents for actions, and workflow agents for coordination.

Quick Start Example

Here's a content creation pipeline that writes, edits, and fact-checks articles:

import { LlmAgent, SequentialAgent } from "@iqai/adk";

// Create specialized agents for each step
const writer = new LlmAgent({
  name: "content-writer",
  model: "gemini-2.5-flash",
  instruction: "Write engaging, informative articles on the given topic.",
});

const editor = new LlmAgent({
  name: "copy-editor",
  model: "gemini-2.5-flash",
  instruction:
    "Edit for clarity, grammar, and style. Maintain the original voice.",
});

const factChecker = new LlmAgent({
  name: "fact-checker",
  model: "gemini-2.5-flash",
  instruction: "Verify claims and flag anything that needs sources.",
  tools: [new WebSearchTool()], // Can use tools for research
});

// Sequential workflow: write → edit → fact-check
const contentPipeline = new SequentialAgent({
  name: "content-pipeline",
  description: "End-to-end content creation with quality controls",
  subAgents: [writer, editor, factChecker],
});

// Use it
const result = await contentPipeline.run({
  message: "Write about the benefits of renewable energy",
});

Choosing the Right Workflow Type

Each workflow type solves different coordination challenges:

Sequential Agents

When to use: Steps must happen in order, each building on the previous result
Perfect for: Content pipelines, data processing chains, quality assurance workflows
Example: Research → Analysis → Report → Review

Parallel Agents

When to use: Tasks are independent and can run simultaneously
Perfect for: Multi-perspective analysis, concurrent data processing, distributed tasks
Example: Sentiment analysis + Topic extraction + Summary generation

Loop Agents

When to use: Need iterative refinement or retry logic until conditions are met
Perfect for: Problem solving, quality improvement, progressive enhancement
Example: Code generation → Testing → Fixes → Repeat until tests pass

LangGraph Agents

When to use: Complex workflows with conditional branching, dynamic routing, or state-dependent decisions
Perfect for: Advanced orchestration, decision trees, adaptive workflows, multi-path processing
Example: Customer support routing → Escalation logic → Specialized handlers based on issue type

Pro Tip: Combine Workflow Types

You can nest workflow agents! Create a sequential pipeline where one step is a parallel workflow, or use loops within larger sequential processes. This gives you powerful composition capabilities.

Common Patterns

Content Creation Pipeline
Sequential: Research → Write → Edit → Fact-check → Publish

Multi-Source Analysis
Parallel: Process documents simultaneously → Aggregate insights

Iterative Problem Solving
Loop: Generate solution → Test → Refine → Repeat until satisfactory

Quality Assurance
Sequential + Parallel: Parallel checks (grammar, facts, style) → Sequential approval workflow

Dynamic Decision Making
LangGraph: Customer inquiry → Route by type → Specialized handlers → Escalation logic → Resolution