Agentic Workflows

An agentic workflow breaks a complex task into stages where AI agents handle each step, with or without human oversight. The key insight: composition beats complexity.

Common Patterns

1. Sequential Pipeline

Each agent's output feeds the next:

Research Agent → Planning Agent → Coding Agent → Review Agent

Best for tasks with clear phases. Each agent is specialized and focused.

2. Orchestrator-Worker

A lead agent delegates subtasks to specialized workers:

# Orchestrator decides what to do
tasks = orchestrator.plan("Build a REST API for user management")

# Workers execute in parallel
for task in tasks:
    worker = spawn_agent(task.agent_type)
    worker.execute(task)

This pattern scales well — add more workers without changing the orchestrator.

3. Evaluator-Optimizer

One agent generates, another evaluates, repeat until quality threshold:

Generator → Evaluator → feedback → Generator → Evaluator → ✓

Useful for code quality, writing, and any task with measurable criteria.

Designing Workflows

When building agentic workflows, consider:

• Checkpoints — save state between stages so you can resume on failure
• Human-in-the-loop — decide where humans approve vs. where agents run autonomously
• Error boundaries — agents will fail; design for graceful recovery
• Cost awareness — each agent turn costs tokens; set budgets and limits

Anti-Patterns

• Over-agenting — using agents for tasks a simple script handles fine
• Unbounded loops — always set maximum iterations
• Shared mutable state — agents stepping on each other's work
• Trust without verify — always validate agent outputs before applying them