Lessons from OpenAI on Building Better AI Agents

The AI Engineer Summit wrapped up a few weeks ago in New York. If you missed it, don’t worry- I’m here to share key insights on AI agents!

One of the most valuable presentations was from Prashant Mital of OpenAI on how to build AI agents. OpenAI has arguably one of the best AI agents in the market today- Deep Research. It is highly capable and clearly ahead of its competitors.

Let’s break down the key lessons from OpenAI on building AI agents effectively.

1. Abstraction is a tool, not a crutch

Avoid using AI agent frameworks at the start of a new project. Writing in primitives (or plain old Python) allows you to familiarize yourself with the workflow and understand its limitations and areas for improvement. Starting with agentic frameworks right away can leave you with a poor understanding of the underlying system and the challenges you may encounter later.

2. Start simple

While there's a lot of excitement about agents becoming co-workers and coordinating them in networks, Prashant emphasizes the importance of starting simple. Begin with a single agent performing a single task. This helps you understand how the agent approaches that specific task, allowing you to refine it until it performs as expected.

3. Use a network of agents for complex tasks

Once you have refined an individual agent, repeat the process until you've mapped out the various components of a complex task. At this point, you’ll be ready to use a network of agents with each specialized in a specific task.

For example, in customer support, you might have:

• One agent focused on conversing with customers
• Another checking the database for relevant actions
• A third handling claims processing

While this design pattern may seem simple, it helps each agent stay focused. AI agents are still limited by context, so assigning them a single type of task enables them to perform at their best.

4. Keep prompts simple, use guardrails for edge cases

It may be tempting to write highly complex prompts, especially for high-stakes tasks like processing refunds. However, this often leads to decreased performance, as agents must juggle multiple aspects at once. Instead, let simple guardrails handle edge cases.

These guardrails could be as straightforward as if-else conditions to ensure valid outputs, or they could involve another LLM for verification.

These insights align closely with my own experience building AI agents- start with primitives, scale gradually, and keep prompts simple. Do they resonate with you as well? Drop a comment below!