Previous post | All posts in series | Next post

Automation tools like If This Than That (IFTTT), Zapier, and Power Automate launched way back in the early 2010s. My favorite automation I’ve built cleans up my calendar by removing canceled meetings and out-of-office notifications. These products simplified what could be previously done through programming: a sequential set of steps, passing along variables and handling conditionals. Some marketers may try to convince you that these types of products and scenarios are AI agents. They aren’t.

Including a language model call in that sequence of steps still doesn’t qualify it as an agent. These automations can be very valuable—imagine how much cleaner my calendar would be if an LLM decided whether a meeting was an out-of-office notice, instead of string-matching on “OOO”! But just having an LLM doesn’t make it an agent.

AI agents should rely on—and do in fact work best with—conversation, the best and worst interface we have ever put on our software.

Disambiguate going in

OpenAI described GPT-3 and later models as having “emergent behavior.” Previous AI models were trained to do a task, and then that task is what they did. Key to the success of ChatGPT is that it tries to follow any instructions typed by the user, not just the task that was trained into it. This is the “magic” of modern chat AI. You can describe literally any task to perform, and the AI will confidently attempt to complete it.

But conversation has value even beyond this. We’ve learned that AI chat is best when we teach it not to attempt to complete the task immediately. It should instead start a conversation by asking you a few questions. This helps you more clearly specify what they want, using the same conversation skills the person grew up learning. Instead of having an AI review your writing after one request, you’ll have better results if the AI makes you answer whether you’re interested in copy editing, a measure of persuasiveness, or some other kind of review. Similarly, it may ask what audience is your writing for, or what are you trying to accomplish?

You could prompt engineer all of these details ahead of time and get excellent results. And you could save that prompt, but what if you next need to write for a slightly different purpose and audience? Having a quick conversation is the fastest way to clearly define what you want to have done.

Let’s consider some more examples where a conversation with an agent will yield much stronger results than an automation:

• An automation could remove canceled and out-of-office calendar meetings, while an agent could help you decide which meetings are important and schedule focus time over the others.
• You could set up an automation to translate your social media content to other platforms and communities. If you used an agent instead, you could describe the latest zeitgeist and trends on a platform and then get tailored results.
• Built-in AI features may suggest you catch up on documents from your manager and direct colleagues. If you work with an AI agent, you can describe what topics you are interested right now and which people’s work you want to catch up on.
• It would be easy enough to set up an automation to summarize all the emails you sent over the week. An agent would ask you which major projects and deliverables are the most important to include.

Conversation is the most powerful and flexible way to create a task for AI, and it’s also the most powerful and flexible way to work with output from AI.

Edit and iterate going out

It’s rare for the first output of an AI to be exactly how you want it. For structured output, perhaps it will work well to use a form over data, but for anything in natural language, conversation is excellent. Is it too long or too short? Do you want to make the tone more formal, or at least lay off the emoji? Just ask, and you’ll quickly get a revision.

The reason this works so well is that humans are natural conversationalists. Meanwhile, AI is naturally good at immediately making revisions to text. You may be tempted to just make the edits yourself, but you’ll often end up spending as much time editing as if you had written it yourself in the first place. If something can be written by AI (and not everything should be), let the AI handle your feedback.

The Sport-Loving Buddy agent I created in the last article gives insights about your local sports teams that you can use to pretend to know about sportsball. But it gets even more useful when you continue in conversation. Here I am forgetting the name of a famous Seahawk:

Let’s consider some examples where the output of agents can be improved through conversation:

• If you’re a US government employee and you need to email DOGE what you did this week, you’ll want to ask the agent to edit in what you know were the most valuable accomplishments.
• AI is an excellent brainstorming partner. After a set of output, tell it which ones you liked and why. Continue to steer the brainstorming until you hit upon the best idea.
• Many people use AI to plan meals and groceries. When I do this, which is itself conversational, I like to follow up by having it order the groceries in “aisle order” so they are easier to pick up.
• Programming is the prototypical example, where even when the AI produces a successful result, it’s worth the time to iterate and improve.

With enough iteration and conversation, any AI slop can be turned into something that is quite good, indeed.

Agents are conversational

Automations or good ol’ fashioned programming remain useful. They are fast, reliable, cheap, and handle plenty of scenarios. Use them whenever you can: there’s a strict order of operations, the task is specifically defined, and the output is consistent. On the other hand, once you recognize the power of AI agents, you’ll be building them centered around conversation for massive productivity and quality gains.

Be sure to subscribe with RSS or on LinkedIn for more in this series on agent best practices!