[Agent Series Volume 1, Part 3 of 3] What is an Agent?

FINALLY!! Finally. Finally!

It’s here. Part 3.

What is an Agent?

In a way the answer has changed over the nearly 7 months since my previous post in this series and, in a way, it hasn’t.

I think a high-level definition of an agent (in the AI/software sense) is “tooling that can do things on a user’s behalf”. I did say ‘high-level’. So….

Anyway, there’s a lot of nuance.

I think my framing in my previous post in this series still applies. Here’s a re-presentation of it with some details added: “show me something” (retrieval/search/results after an action) vs. “do something” (actions performed, sometimes multiple actions performed; display of results not required). I think “display of results not required” is really pertinent. From my growing understanding of agentic use cases, it’s not required that an agent report on what was done. I still think that part is managed by the application or - in more recently popularized terminology - the “harness”.

Thinking of a shopping use case, the agentic steps might be ‘find and purchase’. The non-agentic steps might be ‘define what to find and provide purchase-enabling info’. Again, at a high-level, the user will likely have to provide some level of input to help with context for what they’re looking for and some level of input for information enabling the purchase such as which account to use. Equipped with inputs, the agent can then perform the ‘find and purchase’ steps. On the other side of the process, the user will likely receive some sort of purchase confirmation (e.g. via email) that the agent may or may not have been instructed to report on. From my understanding of agentic flows, technically, and especially if not instructed to provide feedback, an agent can just do the actions and be done.

So I return to my quite high-level definition of an agent - “tooling that can do things on a user’s behalf” - and I’ll add some detail: “with a set of instructions and enablement (e.g. permissions or access)”.


Exemplified

Tooling instructions may vary across application and agentic tooling types. Two examples are included below.

Use Case: Shopping

• Example Application Instruction: show me deals
• Example Agent Instruction: based on deals and the fact that I prefer 50% off clearance, with Google Search, find a purple sweater at the best price and buy it with my CC info → do this on my behalf

Use Case: Job Searching

• Example Application Instruction: show me relevant jobs based on specific inputs
• Example Agent Instruction: start applications for the job(s) I’m likely to be interested in, stop before hitting submit as I’d like to review the company(ies) and opportunity(ies)

Food for Thought…and Conversation

This is a really interesting topic and I’d like to continue the conversation on it. Please feel free to reach out to me if you’d like to too! Also, please stay tuned. I’ve realized that this volume 1 will probably need a bonus post! :D 

Posting again soon,
Jewel