Mastering AI Agents: A Practical Guide to Building Effective Workflows and Tools

Overview

Based on Anthropic’s "Building effective agents", this article provides a detailed practical guide for constructing AI agents and workflow systems, covering when to use agents, design patterns, implementation steps, and tool‑prompt engineering.

When to Use Agents

Agents are suitable for open‑ended problems where the number of steps cannot be predicted in advance, requiring dynamic decision‑making, tool usage, and a feedback loop. They trade higher cost and latency for flexibility.

Design Patterns

Prompt Chain : Decompose a task into ordered LLM calls.

Routing : Classify inputs and direct them to specialized sub‑workflows.

Parallelization : Run independent sub‑tasks or multiple attempts (voting) concurrently.

Orchestrator‑Worker : An orchestrator LLM plans and assigns tasks to worker LLMs.

Eval‑Optimize : One LLM generates output, another evaluates and provides feedback for iterative improvement.

Combined Patterns

Complex applications often combine patterns, e.g., routing + prompt chain for customer‑service queries, or orchestrator‑worker + eval‑optimize for code generation.

Practical Example: Content Moderation

A multi‑step agent evaluates user‑generated text for violence, hate, profanity, political expression, and incitement, applying different thresholds for each dimension and escalating ambiguous cases to human review.

Practical Example: Medical Research Assistant

An orchestrator plans a search strategy across PubMed, agency reports, clinical‑trial registries, and pre‑print servers, distributes the work to specialized workers, aggregates results, and produces a final report with evidence grading.

Implementation Advice

Define clear task scope and success metrics.

Design concise, well‑documented tool interfaces.

Establish feedback loops and supervision checkpoints.

Quantify performance and iterate.

Tool Prompt Engineering

Tools must be described like prompts: precise purpose, input schema, output format, limits, and examples. Prefer simple, token‑efficient formats (Markdown, plain text) over verbose JSON or diff representations.

Example Tool Definition

Planner LLM receives query and produces a search plan:
1. Identify key terms.
2. Choose data sources.
3. Design specific search strategies.
4. Assign workers.
5. Aggregate findings.
6. Determine need for further search.
7. Prepare final report.

Agent‑Computer Interface (ACI) Optimization

Design tools from the model’s perspective, provide examples, clear parameter names, and avoid formats that require exact token counting or complex escaping.

Conclusion

The key takeaway is to start with the simplest prompt solution, evaluate rigorously, and only add agent complexity when it yields measurable improvement. Proper tool design and systematic pattern composition enable reliable, high‑quality AI agents.