Understanding AI Agents
The "Future of AI" Is Already Here
The term "AI agent" has generated significant buzz, but arriving at a precise definition can be challenging in this new and rapidly developing field, which is often dominated by marketing professionals.
To understand what an AI agent is, it's easier to first understand what an AI agent isn't:
Simply asking an AI to complete a task in one go, often called one-shot prompting, is not an AI agent. While you might receive a coherent response, it will likely:
Be vague and not truly aligned with your specific needs, requiring you to re-prompt until you get the answer you want
Not be integrated with your proprietary or personal information, meaning you'll need to upload or integrate it (each time)
Wait until you give it a prompt before providing an answer, requiring you to tell it what to doMost of us use large language models (LLMs) like ChatGPT, Google Gemini, and Claude for generating and editing text, but all we're really doing is making what we did with Google more efficient. While this is valuable, we can do much more.
The Evolution of AI: AI Workflows
A significant step beyond simple prompting is an AI workflow. This is an extension of automation tools such as Zapier or Power Automate. These automation tools take repeatable steps and automate them. To implement them, you must:
Define each step and what actions need to be taken
Connect all the systems you would like to use
Potentially use a "pre-canned" process
All of these require significant thought and planning by you, the human. In our daily lives, we use these for processes like eSignature and sending out follow-up emails/drip campaigns. By incorporating AI models like ChatGPT, we can now:
Ask the LLM a question like "When is my meeting with John?" and it can:
Search my calendar to find it
Check the weather for that day
Convert all that text to a voice note
Play it back to me
While this is still beneficial, the automations are programmed manually. As a user, you must still define the paths and logic, and the only steps removed from a traditional workflow are logical steps. Retrieval Augmented Generation (RAG), where models look up information before answering, is essentially a type of AI workflow. Similarly, NotebookLM functions like an AI workflow – you feed it information, and it can run some predefined operations.
The Evolution of AI: AI Agents – the Present Future
The next level is a truly autonomous AI agent. This is achieved when AI can completely and independently determine the exact steps, select the necessary tools, and go through the circular process of revising and improving its output to accomplish a task.
After receiving a goal, an AI agent:
Uses the LLM to reason and determine the best approach
Takes action using available tools
Observes the interim results
Decides whether further iterations are needed
Produces a final output that meets the initial goal
The most critical difference is that the LLM itself becomes the decision-maker in an AI agent system, replacing the human decision-maker found in workflows.
The potential applications are vast. Imagine a world with:
Research agents (like Perplexity): Process your questions to search, analyze, and synthesize information from multiple sources while maintaining context across conversations
Email management agents: Monitor inboxes, draft responses, and flag important messages for human review
Lead qualification agents: Instantly score leads based on behavior, responses, and engagement, routing high-potential leads to human representatives while filtering or nurturing weaker ones automatically
Productivity agents (like Zivy): Connect various tools such as Teams, Slack, and Jira to manage all your tasks:
Schedule meetings in your calendar
Automatically follow up on low-priority emails
Highlight which emails require your personal review
Automatically decline or reschedule invitations when necessary
There are two approaches to implementing AI agents:
Create the agents yourself – which, while effective, still requires significant planning
Use pre-built solutions – tools like Perplexity or Zivy represent the next evolution in this space