AI has dominated the coding world and ran loops around manual programming once Claude was released. A step above Chat GPT, with a training on excelling on coding tasks, Claude removed manual coding from developers. Instead, coders were told to prompt and design the system that allowed prompts to manage prompts. With Claude Mythos, the one shot with goals and loop increased immensely given the hype it came with. Being banned does indeed generate quite the buzz!
This worked to highlight how efficient machines were with code. With Claude (along with Chat GPT and Gemini) outputting massive amounts of code in mere seconds, the timeline to resolve backlogs shrunk. Think of the prompt, architect a system around it, and voila what was weeks of effort was resolved in minutes.
Problems arose with code maintenance. With humans coding each feature, conflicting logic could be foreseen and a fix adapted. With AI code, these clashes sometimes were just “solved” on the fly.
But now the next level of coding as described by the leaders at the head of the AI frontier are loops. A goal, and then iterate through agents that solve what you are hoping to achieve. Output from different agents are managed until the goal is achieved. In some ways this feels like AGI. High level goal and then iterative attempts at sub goals that are built upon until the goal is achieved.
In fact, like the OpenClaw loops from early in 2026, loops were amazing and then died off. Now ChatGPT and Claude are bringing back the loops with an aspect of moving the human work to a higher level than prompting.
Old way: Iterate through a project with AI and fix things as they come up.
New way: Give a goal, specify different monitoring agents that keep looping through their work and getting sign off until the goal is achieved.
The AI companies have spent $710 billion on AI. VC is currently subsidizing monthly AI plans and even $200 plans do not cover costs. With the focus on agents, the hyperscaleres are bringing about a way to keep token maxing ongoing. However, companies focused on spend and ROI want to know what their cost will be for programmers. Pre-AI you had fixed salary costs, and paid in time as a project went over the deadline.
In the age of AI, a programmer cost can be infinite. A bad prompt can run for hours and consume tokens well above a pre-allocated expected cost for a company. In fact, the AI assisted coding is problematic with fixed spend because when the limit is reached, developers have to do what is now unthinkable; write code manually!
AI coding is here to stay. Goals and loops may not be. The efficiencies generated from the machine writing large chunks of code is too advantageous to give up. The benefit comes at two huge costs for developers, and they are both in the management space. One is the managing token usage, and writing well documented and structured prompts that run the machines efficiently. The second is leveraging the time machine is writing the code. This results in context switching between different parts of the code base and the mental drain on keeping both status in memory.
The AI labs continue to push the capabilities of what is possible with their frontier models. As the developer usage increases, the companies actually making the purchase decisions are now focused not just on the output of the tool, but also the true end to end cost of the tool for its programmers. After all for many companies, their main goal is an efficient bottom line.