AI integration is moving at an astonishing pace. Just a few months ago, we were coming to terms with the idea of AI agents, or what the buzzword mavens call “agentic AI.” Now, we’re starting to look at issues of practical deployment.
If you’re not fully up to speed on agents, that’s okay. Few people are. OpenAI defines agents as “Systems that independently accomplish tasks on your behalf,” with an emphasis on “independently.” ZDNET has a full guide on the topic, which is essential reading.
Also: 8 ways to write better ChatGPT prompts – and get the results you want faster
OpenAI recently released a 34-page PDF entitled “A practical guide to building agents,” which includes notes and guidelines from the company’s experience deploying agents for its clients. After reading the guide, I’ve distilled OpenAI’s recommendations into 10 best practices, detailed below.
(Disclosure: Ziff Davis, ZDNET’s parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)
1. Prioritize stubborn workflows
AI is so dominating the conversation that we sometimes forget to ask an obvious question: when should it be used? We ask that for all our tools. Even if a tool is very general purpose (like a computer or an AI), we still want to know what it’s best at and how it can help.
Also: Is ChatGPT Plus really worth $20 when the free version offers so many premium features?
This is particularly true of agents, which are designed to carry out sets of tasks. When I read OpenAI’s guide, something clicked in terms of my understanding. It was this statement: “As you evaluate where agents can add value, prioritize workflows that have previously resisted automation, especially where traditional methods encounter friction.”
This makes total sense. If you, like me, have been automating workflows for years, you’ve probably run into some that just seem to fight back. They’re either too undefined, too variable, or require too many judgment calls to efficiently automate, even if such automation could save hours or days of time.
Don’t throw agents at every problem you have. If a traditional “deterministic” or algorithmic solution would work, use it. Only move into AI when the old school techniques aren’t up to the task.
2. Understand models, tools, and instructions
Agents consist of three main elements: models, tools, and instructions. The model is the AI itself, the large language model that does the reasoning. Tools are the mechanisms (usually APIs) that agents use to get things done. And instructions are the prompts that are fed to the AIs that instruct the AI agents about what you want and how to get the job done.
Let’s say you run a footwear e-commerce site that lets users upload pictures of their shoes. You want an agent to accept relevant shoe pictures, but filter out all the images that aren’t relevant. The tools are an image reader API and the database API. The model is some LLM that’s used to make a determination.
And the instructions specify that you want footwear, and that you only want to allow footwear that matches the model of shoe being shown. You might also add some guardrails, telling the AI to only accept images from the knee down, only accept images that meet certain standards and practices, only accept images with a certain level of focus and contrast, only accept images without any text, and so on.
Also: Can you build a billion-dollar business with only AI agents (yet)? This author thinks so
By combining models, tools, and instructions, you can automate a task that might have otherwise been difficult to automate. It’s also a task that might have been difficult to assign to human workers because attention spans waver, and human inspection might be slower and less accurate.
3. Build smart, then dumb down
When you see AI models referenced, they often have fairly cryptic names like gpt-4o, gpt-4o mini, gpt-4, gpt-4 turbo, gpt-4.1, gpt-4.1 mini, gpt-4.1 nano, gpt-4.5, o1, o1-mini, o1-pro, o3, o3-mini, o3-mini-high, o3-pro, o4-mini, and o4-mini-high.
Some models are newer than others, and some are less resource-intensive. It turns out that the smarter models tend to require more resources and time. That means they cost more to operate. Models that cost more to operate cost more to rent.
Also: 60% of AI agents work in IT departments – here’s what they do every day
You might think, therefore, that it makes sense to use the least expensive model when you initially field your agents. But OpenAI suggests otherwise. It recommends you assign each task the smartest, most capable model, just to get the task working properly and establish a performance baseline.
Once the task is working, then consider replacing the model with a less capable model. See if the performance stays the same or degrades. I don’t think OpenAI is making this recommendation just to convince you to spend more on AI fees. I think it genuinely wants to make sure your AI solution works before tinkering with cost reduction. That’s the same process that many manufactured products take. Build it, then cost reduce it until you’ve reached an optimal solution.
4. Max out single agents
One of the more intriguing aspects of agents is that you can build teams of agents, much like you would build human teams.
Also: Crawl, then walk, before you run with AI agents, experts recommend
Back when I built the AI Newsroom in 2011, I implemented a fleet of agents, each of which scanned the news for one topic. Then I implemented an editor agent, which accepted the news submissions and validated them as relevant. Another agent grouped related news items together. And yet another agent prepared them for publication on my experimental websites.
Back then, we didn’t have the generative AI tools we have now, so each agent took an entire server, resulting in a couple of racks, to get the job done.
Building the agents was one thing. Getting them all to work together was incredibly difficult. That’s where OpenAI’s recommendation to maximize a single agent’s capabilities first makes so much sense.
Build out slowly. Get one agent to do as much as possible. As soon as you start adding agents, you multiply your complexity.
5. Use prompt templates
One technique OpenAI recommends that can help you max out a single agent is to use prompt templates. These are basically full AI prompts implemented as fill-in-the-blank templates.
This way, rather than switching to a new agent, you just modify the prompt by substituting variables. This technique can prove very helpful not only to reduce agent sprawl but also to make it easier to debug agent behavior.
6. More tools, more agents
OpenAI recommends that, as you add tools, you also add agents. I kind of think of this like building a house. There are different trades involved, including plumbers, electricians, carpenters, roofers, landscapers, and so on. Each of the trades knows their specific skills.
Also: How AI coding agents could destroy open source software
If you’re building up a team of agents, think of each agent as knowing a trade. Assign each agent the tools related to its particular task. As you add new and potentially unrelated tasks, add agents.
7. Tool similarity matters
Apparently, AIs also struggle when there are multiple tools that overlap or have similar functionality. The AIs have difficulty distinguishing when to use which tools.
So, if your tools are very different, a single agent might be able to handle them because it can easily determine what gets assigned to what.
But if you start to assign very similar tools, the agent may skip between the tools. OpenAI recommends, “Use multiple agents if improving tool clarity by providing descriptive names, clear parameters, and detailed descriptions doesn’t improve performance.”
8. Split when agents struggle
This next recommendation is so human, so managerial, it’s almost surreal. Yet it makes sense.
OpenAI says, “When your agents fail to follow complicated instructions or consistently select incorrect tools, you may need to further divide your system and introduce more distinct agents.”
In other words, if your agents become overwhelmed, get them some help. You can tell if they’re overwhelmed if they screw up by not following directions or using the wrong tools.
Also: How to turn ChatGPT into your AI coding power tool – and double your output
If you have an employee or a team who is not following directions or consistently using resources incorrectly, you’d initiate some mediation. Sometimes, it’s time to replace that team member. Other times, it’s simply the fact that your team members have way too much on their plates and need help.
In the case of agents, either split the work among more agents or simplify the process.
9. Build guardrails as layered defense
On a mountain road, guardrails are designed to keep cars that lose control from flying off the edge of a cliff. In AI, guardrails are designed to keep agents that lose control from doing dangerous things.
OpenAI defines seven types of guardrails.
Relevance classifier: Flags off-topic user queries to keep the agent within intended scopeSafety classifier: Detects unsafe inputs like jailbreaks or prompt injection attemptsPersonal information filter: Screens model output for personally identifiable informationModeration: Identifies and blocks harmful content like hate speech and harassmentTool safeguards: Assesses tools based on risk level (for example, what kind of access they have)Rules-based protections: Implements blocklists, regular expression filters, and input limits to prevent known threatsOutput validation: Ensures the model’s responses align with brand or content policies
OpenAI suggests thinking of guardrails as a layered defense, understanding that no one guardrail will provide full protection, but a combined set might help keep things on track.
As with all the other OpenAI recommendations, guardrails should be implemented incrementally. Build in guardrails for risks you’ve identified prior to deployment, then add in and layer on more guardrails when new vulnerabilities are discovered.
10. Plan for human intervention
Face it. Something is going to break. When it does, your guardrails may not catch it. You can’t just set your AIs in action and leave them unattended. AIs are workflow process force multipliers, but they need human supervision.
Also: AI agents make great teammates, but don’t let them code alone – here’s why
Design your system to report to real people. Send anomalous behavior reports to people. Design your AIs to reach out to people when something doesn’t seem right.
OpenAI says that there are two triggers that require human intervention. The first is when an agent or workflow exceeds failure limits. Essentially, when the AI tries and tries and tries and keeps failing. The second is when the AI is about to take a high-risk action, whether that’s doing something that’s not reversible, is sensitive, or could cost a lot of money.
The last thing you want is an AI giving out ginormous refunds without some level of human oversight. AIs are perfect targets for malicious intervention. We humans aren’t perfect, but we can often spot things that our AIs can’t.
Best practices
OpenAI recommends that no matter how you’re planning on building an agentic AI solution, keep the components flexible and drive them by clear, well-structured prompts. Start small, validate with real users, and grow capabilities over time.
If you think about it, these are the same guidelines IT installations have used since the early days. The more things change, the more they stay the same.
Have you started experimenting with AI agents in your workflows? What kinds of tasks have proven hardest to automate? Do you think agents could finally make them manageable? Which of OpenAI’s best practices, like prompt templates or layered guardrails, resonate most with your own experience? Are there areas where you think human oversight is still absolutely essential? Let us know in the comments below.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.
Get the morning’s top stories in your inbox each day with our Tech Today newsletter.
