The world of AI is buzzing with a new term: agents. If you’re in the tech space, you’ve likely seen the flood of demos, open-source projects, and think pieces. The promise is incredible: AI that doesn’t just respond, but acts. It’s a significant leap, moving from passive tools to active partners. As an engineer who has been in the trenches of AI development for years, I’ve seen this excitement before. But I’ve also seen the #1 mistake that brilliant teams make: building a technically impressive agent that does… well, nothing of value. They fall in love with the how and completely forget the why. This series is your guide to avoiding that trap. We’re going to build an AI agent from the ground up, but we’re going to do it the right way. And it all starts with the most critical, yet often overlooked, phase: planning. Before you write a single line of code, you need to give your agent a mission.
What is an “Agentic” AI Solution, Anyway?
Let’s demystify the term. At its core, an AI agent is a system that uses an AI model (its “brain,” typically a Large Language Model like GPT-4) to interact with its environment to achieve a specific goal.
Think of it like this: if you ask a standard AI model, “What’s the weather in London?” it will give you text-based information it was trained on. If you ask an agent, it can actually go out, use a weather API (a “tool”), and give you the real-time forecast. The key difference is the ability to take action.
The “AI Agents and LLMs Report” uses a great analogy: Alfred the Butler. When you ask Alfred for coffee, he doesn’t just describe what coffee is. He understands your request, formulates a plan (go to the kitchen, use the coffee machine), uses his tools (the machine), and completes the objective (brings you the coffee). That’s an agent.
Do you need Agentic AI at all?
Before you dive headfirst into building a complex agent, it’s crucial to ask the most important question: is an agent even the right tool for the job?
Agentic AI implies you give autonomy for the AI to choose its next action. This is incredibly powerful, but it also opens up the possibility of getting undesired results, such as the agent getting stuck in a loop, choosing the wrong tool, or hallucinating a disastrous sequence of actions. More often than not, what you want to achieve can be done with a more controlled, fixed workflow that only needs the brainpower of an LLM for specific steps. Forcing an agentic solution where it’s not needed is an overkill at best, and a total disaster at worst.
To figure out if you truly need an agent, ask yourself these questions about the task you’re trying to automate:
- Is the path to the solution fixed or dynamic? If the task always follows the same sequence of steps (e.g., Step A, then B, then C), you don’t need an agent. You need a script or a simple application that calls an LLM when necessary. An agent is only needed when the next step is uncertain and depends on the outcome of the previous one.
- How complex is the decision-making? Does the system need to weigh multiple options, adapt to new information in real-time, and make strategic choices? If so, an agent might be a good fit. If the “decision” is just “if this, then that,” a traditional program is more reliable.
- What is the cost of failure? If an agent making a mistake could result in sending an incorrect invoice, deleting the wrong file, or giving a customer bad information, the risk of full autonomy is high. In these cases, a “human-in-the-loop” system or a rigid workflow is a much safer bet.
Many powerful AI solutions are not agentic. They are well-designed systems that use LLMs for what they do best—understanding and generating language—within a predictable and reliable structure. Don’t chase the “agent” buzzword if a simpler, safer solution will solve your problem more effectively.
The above image displays a simple thought process to help you determine if Agent AI is what you need. With a task in mind, ask yourself
-
Is the path to the solution dynamic and uncertain? If the answer is no, all you need is a fixed automation workflow.
-
Does the task require adapting to new information in real-time? If the answer is no, you may just need a fixed automation workflow with access to real-time values via, say API call, and some brain power from the LLM to process it slightly.
-
Otherwise, you may be handling a task that is dynamic, requiring the LLM to decide which tools or real-time information to use depending on certain triggers (such as from a user). This is where you need Agentic AI.
When a Fixed Workflow is Better
Here are a few scenarios where a predictable, non-agentic workflow is the superior choice:
- Daily Sales Reporting: Your task is to email a sales report every morning at 9 AM. The steps are always the same: 1. Query the sales database for yesterday’s numbers. 2. Format the data into a table. 3. Email the table to the sales team, perhaps using an LLM to write a nice summary of the key takeaways. This is a perfect job for a simple, scheduled script. An agent’s autonomy is unnecessary and adds needless complexity.
- Invoice Processing: You need to extract the vendor name, date, and total amount from PDF invoices and enter them into an accounting system. While an LLM can be used for the extraction step, the overall workflow is fixed: 1. Receive PDF. 2. Call LLM to extract data. 3. Insert extracted data into the accounting API. There’s no complex decision-making required.
- Content Summarization: A user pastes a long article and wants a summary. The workflow is linear: 1. Take user’s text. 2. Send it to an LLM with a prompt like “Summarize this article in three bullet points.” 3. Display the result. An agent trying to “decide” how to summarize would be overkill.
In all these cases, the path is clear and the goal is well-defined. A simple, reliable workflow is more efficient and safer than a fully autonomous agent.
The North Star: Why a Clear Mission is Everything
Once you’ve determined an agent is the right approach, you need to give it a mission. Without a clear objective, your agent is a ship without a rudder. It might have the most powerful engine in the world, but it will drift aimlessly, a recipe for wasted time, wasted resources, and a final product that, while technically clever, is practically useless.
Your agent needs a well-defined North Star, which mainly involves:
- The Agent’s Identity: The mission defines the agent’s persona and purpose. Is it a helpful customer service assistant? A ruthless data analyst? A creative marketing copywriter? This identity, which you’ll later encode in a system prompt, governs its tone, behavior, and the boundaries it operates within.
- The Agent’s Toolkit: The mission determines what tools the agent needs to do its job. An agent designed to book flights needs access to airline APIs and a calendar, not a code interpreter. An agent built for market research needs web search and data analysis tools. The mission clarifies what “hands and eyes” your agent needs.
Simple, High-Impact Examples
Let’s put it all together with a couple of simple, high-impact use cases where an agent’s flexibility shines.
Use Case 1: The Customer Support Agent for Reservation of Flight & Hotels
A user doesn’t say, “Search flights from JFK to LAX on December 5th.” They say, “I need to get to LA for a wedding next month. I want to fly out on a Friday after work and come back Sunday night. Find me a cheap flight and a decent hotel near Santa Monica.”
- Why an Agent is Better: A fixed workflow would fail instantly. It can’t interpret “next month,” “after work,” or “decent hotel.” An agent, however, can dynamically plan and execute. It can use a calendar tool to figure out the dates, a flight API with a time filter, a hotel API with a location filter, and even a review site API to determine what a “decent” hotel is. It can handle follow-up questions like, “Are there any cheaper options if I fly back on Monday morning?” without starting over. The path to the solution is conversational and changes with each piece of new information.
Use Case 2: Data Analyst Agent
A manager doesn’t ask, “Run sales_report_Q3.sql.” They ask, “Why were our sales in the Midwest flat last quarter while the West Coast grew by 15%? Look into our marketing campaigns and see if there’s a connection.”
- Why an Agent is Better: This is an investigative task, not a fixed report. A rigid script can’t form a hypothesis. An agent can reason: “First, I need to confirm the sales numbers from the sales database. Then, I’ll pull marketing spend for both regions from the marketing database. Then, I’ll use a data analysis tool to see if there’s a correlation. Finally, I’ll search the web for competitor activity in the Midwest during that period.” The agent chooses its tools and sequence of actions on the fly, driven by an open-ended question.
Use Case 3: Deep Research Agent
A product manager doesn’t ask, “Summarize the Wikipedia page for ‘Machine Learning’.” They ask, “Create a report on the current state of generative AI in healthcare, including the top startups, the latest regulatory challenges, and the most promising use cases for diagnostics.”
- Why an Agent is Better: This is an exploratory process of discovery. A fixed workflow can’t dig deeper. An agent can start with a broad search, identify key entities (companies, regulations, technologies) from the initial results, and then launch new, more targeted searches for each of those entities. It can decide to search academic papers for use cases, news articles for startups, and government websites for regulations. The agent builds a complex understanding by iteratively exploring different avenues, much like a human researcher would. The workflow is not only dynamic but self-directed.
The takeaway is simple: agents excel at navigating uncertainty. They thrive on the messy, open-ended problems where a fixed instruction manual would fail. Your job, in this planning stage, is to give them a clear destination. The agent will figure out the path.
This planning stage isn’t the most glamorous part of building an AI agent, but it is the most important. A well-defined mission is the foundation upon which everything else is built.
In our next post, we’ll talk about what constitutes a functional AI agent. Subscribe to get it in your inbox!
FYI: More articles can be found in our dedicated Blogs on Substack.