AI Agents > Single Questions
One of the biggest problems with Chatbots is that they balk when you ask them a tough question. One way to expand? Use AI Agents. Read: "Everything Starts Out Looking Like a Toy" #224
Hi, I’m Greg 👋! I write weekly product essays, including system “handshakes”, the expectations for workflow, and the jobs to be done for data. What is Data Operations? was the first post in the series.
This week’s toy: a wooden satellite was launched into space this week. It’s an important test for the viability of mass timber for situtions where steel, aluminum or composites would normally be used. Edition 224 of this newsletter is here - it’s November 11, 2024.
Are you using AI agents or thinking about this problem? Drop an email!
The Big Idea
A short long-form essay about data things
⚙️ AI Agents > Single Questions
“Hey ChatGPT, can you do a few things for me? First, I want you to search Google for some information, then collate it into a list, and then check that list to make sure it’s correct. Return it to me as a .csv file and set the columns according to my directions: company, website, address, phone number, and site summary. Highlight the companies that are most likely to buy.”
I'm happy to help you create a CSV file template with the columns you specified: company, website, address, phone number, and site summary. However, I don't have the ability to perform live Google searches to gather current information. If you'd like, I can provide a sample CSV file with placeholder data or guide you on how to collect and organize this information. Please let me know how you'd like to proceed.
If you carefully phrase your prompt above to do account research, you can get ChatGPT, Claude, or the LLM of your choice to complete the search. You’ll also likely reach a point where these services balk and refuse to create lists or continue to do repetitive work for you.
Getting one result right is very different than doing this multiple times.
It’s all about the context
Why does ChatGPT refuse to search this way? There are a few obvious reasons:
At the beginning of your search, it doesn’t have enough context to provide better results
it’s expensive to create lists based on multi-step processes when you’re not charging by the token bucket
the more operations it completes, the more likely it will be to wander
Adding context is key here. You’ll need an agent-style architecture to move from one-shot AI questions to a process run in multiple steps. What’s that, exactly?
It means separating the component steps of your workflow into separate actions handled by different parts of an AI workflow.
Enter the Agents
Inspired by Crew.ai’s concept of an Agent “Crew”, a group of AI agents addresses these signature deficits in the following way:
adds context by explaining what you want from each step
limits token usage to lists that have already been created
asks the LLM to do less at a time to make it easier to check and harder to wander
In this model, an agent has a role, a goal, and a backstory. A role might be “data analyst” or “sales development representative”. A goal is ideally a concrete step, like “find 10 accounts per session,” and a backstory is the context you need to give to the LLM about your process and company.
The Account Discovery Agent is designed to assist Business Development Representatives (BDRs) by automating the search and identification of potential client accounts that align with the company's target market. The agent streamlines the initial stages of the sales pipeline by providing curated lists of high-potential accounts, allowing BDRs to focus on outreach and relationship building.
Once you establish these parameters for an agent, you let it know which tools it can use, what guardrails it should have, and what output it should return. Optionally, you can have the agent delegate work or end when they complete their task.
Changing from one-time to continuous
Adding an AI agent to your workflow is the mental step to move from one-time actions to continuous refining. Whether you are creating account lists or doing other forms of data discovery, agents let you break these items into individual steps. Creating individual personas for these agents insulates you from some of the information drift that LLMs create.
But “continuous” also means “continuous review by a human”. Automated process needs a good deal of dialing in to be right for your business. Building tests that let you know when things are not working helps you take advantage of the productivity AI Agents can provide.
What’s the takeaway? AI agents expand your ability to use LLMs by making it possible to break down the tasks you’re doing into their component parts. Improving each step of your process makes it a lot easier to scale it.
Links for Reading and Sharing
These are links that caught my 👀
1/ Everything you wanted to learn about lines but didn’t know to ask - this is a solid introduction to operations science, a.k.a. why is that line so long?
2/ Sharing to be understood - When you work remotely, it’s often hard to know how much to share about current progress to keep others informed. Wes Kao’s essay on providing the right amount of context is a great reminder of what to include in your tl;dr.
3/ Read a little bit more than you do today - The average American reads 16 minutes a day. Read more than that and you’ll expand your brain. (No, not on your phone either. It has a much better effect if you pick up a physical book and turn the pages.)
What to do next
Hit reply if you’ve got links to share, data stories, or want to say hello.
The next big thing always starts out being dismissed as a “toy.” - Chris Dixon