Building Effective Agents

When building an LLM agent, you should start with the simplest solution rather than raising complexity from the outset, and add structure only when it actually improves performance. Anthropic distinguishes between workflows and agents, recommending workflows for predictable tasks and agents only when flexible judgment and long-running execution are required.

The basic building block is the augmented LLM. The core idea is an LLM equipped with retrieval, tools, and memory that generates its own queries, chooses the appropriate tool, and retains the information it needs. This implementation can be connected more cleanly through a standardized interface like the Model Context Protocol, but what matters most is tailoring and documenting the capabilities for the specific use case.

Commonly used patterns in practice are as follows.

• Prompt chaining: A structure where the output of one step becomes the input of the next, suitable when the task can be cleanly decomposed into fixed subtasks.
• Routing: Classifies input and sends it to different prompts, tools, or models.
• Parallelization: Runs independent subtasks simultaneously, or runs the same task multiple times and aggregates the results.
• Orchestrator-workers: A central LLM dynamically breaks down the needed subtasks and distributes them to multiple workers.
• Evaluator-optimizer: One model produces an answer while another repeats evaluation and feedback to improve quality.

Agents are more autonomous than these workflows. The LLM plans and uses tools on its own while observing environmental feedback, and only asks for human confirmation when necessary. However, since costs and the risk of accumulating errors grow larger, it's important to test thoroughly in a sandbox environment and have clear tool documentation and a reliable agent-computer interface (ACI).

Ultimately, the standard for success is not the most sophisticated system but the system that fits the problem. You should first check whether a simple prompt and evaluation can solve it, and only expand to a multi-step workflow or agent when that proves insufficient.