Anthropic Agentic Systems: A Five-Part Exploration is sponsored by Agent.ai - Discover, connect with and hire AI agents to do useful things.
Last month Anthropic came out with their take on Agentic Systems. Among all the hype about Autonomous Agents, I like this write up as it addressed 2 approaches -
Workflows - systems where LLMs and tools are orchestrated through predefined code paths.
Agents - systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.
I strongly believe that in reality we will have Agentic systems that not only fall into the above 2 categories but are also a hybrid whereby some tasks are pre-defined and some tasks are autonomous. We will have a whole spectrum of Agentic Systems.
To make Agentic Systems real to people Anthropic identified 5 common patterns they've seen in production. The intent of this Five-Part exploration is to do a deep dive on each of them. Starting with the simplest to the more complicated ones and put a low code tools like Agent.ai to the test.
While Anthropic had written their paper targeted at developers, I have been channelling my own advice that Everyone is a programmer and will implement these patterns and demystify what seems like complicated concepts, with simple examples we can all understand.
From Anthropic - In this post, we share what we’ve learned from working with our customers and building agents ourselves, and give practical advice for developers on building effective agents. https://www.anthropic.com/research/building-effective-agents
So this is my first of 5 posts, where I will explain the concept,give you an agent you can play with that demonstrates this concept and share a video showing how the agent was created.
There is one part of the paper that I am not sure I agree. Anthropic talks about the Augmented LLM
The basic building block of agentic systems is an LLM enhanced with augmentations such as retrieval, tools, and memory. Our current models can actively use these capabilities—generating their own search queries, selecting appropriate tools, and determining what information to retain.
In this visual, the LLM is orchestrating everything. So if I need to call a tool to do a Web Search, it is the LLM that makes the call. While I understand why the LLM vendors would like this (more tokens), I think there are many use cases where it is cheaper/faster/deterministic to just run regular code. Agent.ai does do this and includes a combination of Augmented LLM functionality and standalone tools (as you will see in this example). It remains to be seen if there is a right or wrong way.
The agent that I will use to showcase each of these patterns takes in the name and company of a person and uses their LinkedIn profile to generate a SuperHero bio for them.
Here is a link to the agent if you want to try it out - Prompt Chaining - Your SuperHero Story
So what is Prompt Chaining
Prompt chaining is a technique used in GenAI where multiple prompts are linked together in a sequence to achieve a more complex or refined output. The output of one prompt serves as the input for the next, creating a "chain" of interactions that build upon each other to produce a more sophisticated or nuanced result.
A simpler way to think about it is that it is like a recipe that you follow from one step to the next. Here is a visual from Anthropic.
So if we come back to our SuperHero bio generator, here are the prompts that we would chain in serial -
Get the name and company of the person (user provides this input)
Using the name and company as input, search the web for their LinkedIn URL
From the web search results extract the LinkedIn URL
Using the LinkedIn URL connect to LinkedIn to get their profile information
Using the profile passed, read their profile and generate their superhero story.
Output their story.
The process looks like this (output of step 1 is input to step 2 and so on) -
If we were to look at these steps in Agent.ai, this is how that looks -
and here is the output - cringy as would be expected.. Vikram "The GTM Alchemist" Ekambaram, does have an interesting sound. I sometimes do feel like a mad scientist trying to make gold for my clients with agents!!!
Join Vikram on his journey as he continues to break barriers and forge new paths in the ever-evolving world of technology and business strategy.
If anyone wants to see how I built this agent showcasing prompt chaining here is a link to the video.
