Peter Steinberger recently shared a line that crystallized something I have been thinking about:

“You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.”

The first phase of working with LLMs was largely about prompting. How do you ask the right question? How do you provide enough context? How do you structure the instruction so the model gives you something useful?

Then the work evolved. We started using multiple agents, sub-agents, tool calls, retrieval, code execution, evaluations, and more structured workflows. The model became less of a chat interface and more of a component inside a broader system.

But the next step is about designing loops. A good loop can plan, act, inspect, critique, revise, and continue. It can generate work, test that work, identify gaps, update its approach, and try again. It can route tasks to specialized agents. It can decide when more context is needed. It can compare outputs against a standard. It can surface failure modes. It can improve the next iteration.

That is very different from sitting in front of a model and manually prompting every step.

It is also an area where I am not yet where I want to be.

I have become much more effective at using LLMs in my day-to-day work. I use them to accelerate analysis, coding, writing, synthesis, brainstorming, and structured thinking. I have also gotten better at decomposing larger tasks across multiple agents or sub-agents. But I still remain in the critical path more often than I would like.

Too often, I am still the orchestrator. I decide the next step. I inspect the output. I create the next prompt. I ask for the revision. I move the work forward. That can be very productive, but it does not fully capture the leverage this technology is starting to make possible.

The frontier is moving from prompt craft to system design.

The question is no longer just, “How do I ask the model the right thing?” Increasingly, the better questions are:

• How do I design agent loops that can make progress without constant intervention?
• How do I build in critique, verification, and evaluation?
• How do I preserve quality and direction while removing myself as the bottleneck?
• How do I give these systems the right goals, constraints, context, and stopping conditions?
• How do I make the loop trustworthy enough to use for real work?

This feels like a good learning goal for June and July.

One of the most exciting and humbling parts of working in technology right now is how quickly “advanced” becomes “basic.” Practices that felt sophisticated a few months ago can suddenly feel like the starting point. That requires a different posture: more experimentation, more humility, more willingness to revisit assumptions, and more comfort being a beginner again.

That is the part I want to keep leaning into.

I do not just want to use the current tools better. I want to understand the new workflows they make possible. I want to keep pushing on what technology enables us to do, how fast we can learn, and how quickly we can convert new capabilities into better systems.

For me, that means moving from being the person who prompts every step to the person who designs the system, defines the quality bar, creates the feedback loops, and knows when to intervene.

The next leap in productivity will not come only from better prompts.

It will come from designing better loops.

---

This writing reflects my personal perspectives on product management, AI, and content discovery. It does not represent the official position of my employer or any affiliated organization.