الهندسة في أنثروبيك: دليل عملي لبناء وكلاء نماذج لغوية كبيرة فعالة

In a recent blog post, Anthropic shared their insights on building effective Large Language Model (LLM) agents, drawing from their experience working with numerous teams across various industries. The key takeaway is that simplicity and composability trump complex frameworks when it comes to successful agent implementation. This post breaks down Anthropic’s approach to agentic systems, offering practical advice for developers looking to leverage the power of LLMs.

Defining Agents: Workflows vs. Autonomous Agents

Anthropic distinguishes between two types of agentic systems: workflows and agents. Understanding this distinction is crucial for choosing the right approach for your application.

• Workflows: These are systems where LLMs and tools are orchestrated through predefined code paths. They offer predictability and consistency for well-defined tasks.

• Agents: These are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks. They provide flexibility and model-driven decision-making at scale.

When to Use (and Not Use) Agents

The post emphasizes the importance of starting with the simplest solution possible. Agentic systems introduce complexity and often trade latency and cost for improved task performance.

• Consider whether the added complexity is truly necessary. Optimizing single LLM calls with retrieval and in-context examples might be sufficient for many applications.
• Workflows are suitable for well-defined tasks requiring predictability.
• Agents are the better option when flexibility and model-driven decision-making are needed.

Frameworks: Use with Caution

While frameworks like LangGraph, Amazon Bedrock’s AI Agent framework, Rivet, and Vellum can simplify initial implementation, they can also introduce unnecessary abstraction and complexity.

• Frameworks simplify tasks like calling LLMs, defining tools, and chaining calls.
• They can obscure the underlying prompts and responses, making debugging more difficult.
• Start with LLM APIs directly to understand the underlying code.
• Incorrect assumptions about the framework’s inner workings are a common source of error.

Building Blocks, Workflows, and Agents: A Progressive Approach

Anthropic outlines a progressive approach to building agentic systems, starting with the foundational building block and gradually increasing complexity.

• Augmented LLM: The basic building block is an LLM enhanced with augmentations such as retrieval, tools, and memory. Focus on tailoring these capabilities to your use case and providing a well-documented interface.

• Prompt Chaining: Decompose a task into a sequence of steps, where each LLM call processes the output of the previous one. Use programmatic checks (“gates”) to ensure the process stays on track. Ideal for tasks easily decomposed into fixed subtasks.

• Routing: Classify an input and direct it to a specialized followup task. Useful for complex tasks with distinct categories handled separately.

• Parallelization: Run LLMs simultaneously on a task and aggregate their outputs programmatically. This manifests in two variations: sectioning (breaking a task into independent subtasks) and voting (running the same task multiple times for diverse outputs).

• Orchestrator-Workers: A central LLM dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results. Well-suited for complex tasks where subtasks aren’t predefined.

• Evaluator-Optimizer: One LLM generates a response while another provides evaluation and feedback in a loop. Effective when clear evaluation criteria exist and iterative refinement provides measurable value.

• Autonomous Agents: Agents plan and operate independently, potentially returning to the human for further information or judgement. Crucial to gain “ground truth” from the environment at each step and include stopping conditions. Suitable for open-ended problems where the required steps are unpredictable.

Practical Applications

The post highlights the practical value of AI agents in customer support and coding, emphasizing that agents are most valuable for tasks that require conversation and action, have clear success criteria, enable feedback loops, and integrate meaningful human oversight.

• Customer Support: Combines chatbot interfaces with tool integration to access customer data, order history, and knowledge base articles, enabling programmatic actions like issuing refunds.
• Coding Agents: Leverage automated testing for verification, enabling iterative solution refinement based on test results. Human review remains crucial for broader system alignment.

Prompt Engineering Your Tools

The importance of well-defined tools and documentation is emphasized. Investing in a good agent-computer interface (ACI) is as crucial as investing in human-computer interfaces (HCI).

• Give the model enough tokens to “think” before writing itself into a corner.
• Keep the format close to what the model has seen naturally occurring in text.
• Avoid formatting “overhead” such as keeping accurate line counts or string escaping.
• Test how the model uses the tools and iterate.
• Change arguments to prevent mistakes (Poka-yoke).

Conclusion

Anthropic’s core message is clear: prioritize simplicity, transparency, and a well-crafted agent-computer interface. Start with simple prompts, optimize them, and add multi-step agentic systems only when simpler solutions fall short. By following these principles, you can build powerful, reliable, and trusted agents.

Source: Anthropic