AI Agent Patterns

Design patterns for building LLM-powered agents with durable execution, parallel processing, and human oversight.

Agent Architectures

Orchestrator-Worker

A central orchestrator dispatches work to specialized worker agents. Each worker handles one domain (research, writing, analysis) and returns structured results.

Evaluator-Optimizer

An evaluator scores output quality and loops back to the generator with feedback until quality thresholds are met.

Router

Classifies incoming requests and routes to the appropriate specialized handler. Useful for multi-intent systems.

Key Patterns

Parallel Workers

Run multiple workers simultaneously for independent sub-tasks. Collect and merge results when all complete.

Human-in-the-Loop

Pause execution at critical decision points and request human approval before proceeding. Essential for high-stakes operations.

Tool Calling

Equip agents with tools (database queries, API calls, calculations) and let the LLM decide when to use them based on the task.

Best Practices

• Use structured output schemas for all LLM calls
• Implement retry logic with exponential backoff for external API calls
• Set hard timeouts to prevent runaway agent loops
• Log every decision point for debugging and audit