OpenAI’s guide cuts through the hype. Agents are not magic. They are loops that call tools until a task is complete.

The core: give a model access to tools, let it decide which tool to use, execute the tool, return the result to the model, repeat until done. That is an agent.

Where it gets interesting is orchestration. When do you give the agent freedom and when do you enforce structure? Too much freedom and the model hallucinates paths that lead nowhere. Too much structure and you might as well have written a script.

The practical lesson: start small. One tool, one task, clear boundaries. Add complexity once you understand where things go wrong. Most agent projects fail not from too few capabilities, but from too many.

Werner Vogels wrote seven laws for cost-conscious architecture. The core: cost is an architecture decision, not an afterthought.

The first law is the most important. Make cost a non-functional requirement. Just as you factor in performance and security during design, factor in cost. Not when the bill arrives, but upfront.

The second law: observe and measure what you spend. You can’t optimise what you don’t measure. This sounds obvious, but most teams don’t know what their system costs per transaction, per user, per feature.

The lesson is broader than cloud architecture. Every system you build has a cost side. Ignore it and you discover that too late.

Most managers focus on what goes wrong. Understandable. Problems demand attention. But the effect is that people only hear from you when something is wrong.

Sinek flips it. Catch people doing things well. Not waiting for the annual review, but in the moment. “You did this well, and here is why it matters.” It inspires the person and signals to the team what good behaviour looks like.

On feedback: we tend to give it the way we ourselves want to receive it. That is the mistake. The question is how the other person wants to hear it. Some people want directness. Others need time to prepare. “Is this a good moment?” or a scheduled meeting can be the difference between feedback that lands and feedback that bounces off.

Timing too. Unless something is urgent or dangerous, save improvement points until the end of the day. Let someone focus on the work without being interrupted by criticism. Later there is space for reflection.

The bigger point: showing interest in who someone wants to become, not just what they do for you. What do they dream of? What do they want to achieve, even outside work? The goal is not just that someone performs. It is that someone leaves the organisation as a better version of themselves, because they felt supported in their growth.

The question is not whether you hit your targets. It is whether the people who worked for you became better for it.