Building Effective Agents

The most successful agent implementations use simple, composable patterns rather than complex frameworks. Success comes from building the right system for specific needs, not the most sophisticated one.

Agentic Systems Taxonomy

The article distinguishes two categories:

Workflows: LLMs and tools orchestrated through predefined code paths
Agents: LLMs dynamically direct their own processes and tool usage

Start with workflows when possible. Move to agents only when you need flexibility at runtime.

Workflow Patterns

Five composable patterns for orchestrating LLMs:

Prompt Chaining — Sequential decomposition with gates between steps. Each LLM call handles one focused task.
Routing — Classification directs tasks to specialized handlers. Good for separating distinct request types.
Parallelization — Run multiple LLM calls concurrently. Use "sectioning" to split work or "voting" for consensus.
Orchestrator-Workers — Central LLM delegates subtasks dynamically. Useful when work can't be predetermined.
Evaluator-Optimizer — Iterative refinement loops where one LLM generates and another critiques.

Core Principles

Three principles guide effective agent design:

Simplicity first. Resist adding complexity. Each abstraction layer adds latency and failure modes.

Transparency in planning. Make the agent's reasoning visible. Users need to understand what it's doing and why.

Agent-Computer Interface (ACI). The interface between agent and tools matters as much as the UI between human and software. Invest in clear documentation, error messages, and tool design.

The Building Block

Every agentic system builds on the augmented LLM: a language model enhanced with retrieval, tools, and memory. Get this foundation right before adding orchestration.

Connections

Extends the design patterns covered in agentic-design-patterns. Complements build-autonomous-agents-using-prompt-chaining-with-ai-primitives with Anthropic's production insights. See 12-factor-agents for operational best practices.