The 3 Rules Anthropic Uses to Build Effective Agents

Hey 👋 

Hope you’re having a great day!

Yesterday I’ve posted a deep dive post on Reddit: The 3 Rules Anthropic Uses to Build Effective Agents, if you’ve not checked it yet, I highly recommend to see it. Beside of that I’ll post a bonus content on a same topic here.

Here’s the original post on Reddit 👇️ 

As you know Anthropic is a company behind a Claude, one of the most popular LLMs on the market, that’s specifically great at generating a code. So it worth listening to them, what they suggest to keep in mind during building agents.

To be honest they don’t suggest something very clever that you or me haven’t thought about before, but it’s great mental model to keep while building agentic systems.

Three core rules that we should keep in mind while building agentic systems are:

• Don't build agents for everything. That’s what I always tell people. Have a filter for when to use agentic systems, as it's not a silver bullet to build everything with.

• Keep it simple. That’s the key part from my experience as well. Overcomplicated agents are hard to debug, they hallucinate more, and you should keep tools as minimal as possible. If you add tons of tools to an agent, it just gets more confused and provides worse output.

• Think like your agent. Building agents requires more than just engineering skills. When you're building an agent, you should think like a manager. If I were that person/agent doing that job, what would I do to provide maximum value for the task I’ve been assigned?

You know me already, you know that I always tell people: don’t build agents or even agentic systems for everything. I know it’s sometimes very tempting to use the latest technologies to build something with them and yes, you should definitely build whatever comes to mind, but test it, think it through, and ask yourself “Is this the right choice of technology for my project?”
Build a couple of projects with agentic systems and you'll start to develop that mental model. I guarantee it.

I’m just gonna skip the second point, as it’s quite clear why we need to keep it simple: lower cost, easier debugging, fewer errors, easier to scale and much more.

The third point is actually extremely interesting, and I think that’s key takeaway from this post. Imagine being the agent and the prompt you give to the agent is given to you. Is it enough to do the job? Does it have the clarity you need? Do you have all the tools necessary to accomplish the task?

I’ve used that mental framework, and it helped me a lot. Imagine your manager tells you to build a website, you'd have tons of questions. What should it do? Who’s the target audience? Do we need authentication, a blog, a forum, etc.? You’d go back to your manager and ask for more clarity.

Now imagine you’re an agent and you can’t ask for more clarity. You need to have everything in the prompt to achieve the desired outcome, otherwise you’ll improvise.

Try giving a step-by-step guide, try giving the agent more information about your desired result. I bet it will perform 10x better.

Here’s the original post from Anthropic about this mental model.

Hope you like it!

Yes, it wasn’t planned for me to send an email today, but I thought, this is a rapidly developing world, especially when it comes to AI and AI agents. So I wanted to share this with you today, so that tomorrow you’ll be a bit more experienced than you were yesterday.

Have a great rest of the day 👋 

George