Building agents

🕵🏻‍♀️ Agentic design patterns

Agents are slowly becoming necessary tools when building AI solutions but what exactly constitutes an agent? While an exact definition is still up for debate one useful distinction that is worth making is between

⚙️ agentic workflows that have a static execution path predetermined from the creator

🤖 and those with a more dynamic execution path determined by the AI

Agentic workflows are for the time being more useful for industry applications because of their modularity, auditability and the ability to test their behavior better. Here are some common agentic workflows from Antropic that we also have encountered in projects

🔗 Chaining e.g. producing an answer, verifying it along some criteria

🔀 Routing e.g. sending a customer request to the appropriate department or specialised LLM

🤔 Reflection e.g. produce an answer and reflect on the correctness until the answer is correct

Read more in this very useful post https://www.anthropic.com/research/building-effective-agents