I never went to agentic loops. I've been doing "agents" since the initial release of OpenAI's API. I've also been an active AI researcher for 40 years, surviving the winters by also being a 3D graphics and animation pipeline developer.

I am approaching AI as not an automation technology, but as a cognitive assistance technology. My AI Agents are all "method actor prompted" and placed inside other open source software, where they are created to believe they are one of the developers of that open source application. That creates "intelligent" versions of those applications.

For example, a word processor I have integrated agents into that are able to support editing and layout of the contents of the editor, but more important the user can create solo and groups of agents that have knowledge of the subject document, and understand how to communicate better than the user - BUT they do not write for the user, they coach the user to write and communicate better.

Are you familiar with Chain-of-Thought / Reasoning models? I've made a fully transparent system that is similar I call "Chain of Method Actors" that operates very much like Chain-of-Thought models, but the user is included in the loops, in a conversational exchange between their experts and the user. Net result? Massive efficiency and expense savings. Where CoT models may spin in their agentic loops hundreds of times before they emit their best guess what they user wants, my Socratic Chain of Method Acting Agents will begin with clarifying questions, rather than guessing, and the eventual answer the user seeks is often delivered in 3 to 6 loops. Not hundreds.

But my goal here is not the same as most: my system I have designed is for people doing critical work they must understand and stand behind. This is a knowledge IDE that I have created. It's an office suite with conversationally programmable AI Agents that method act the expertise that one needs. It is not automation, it's human enhancement, intellectual cognitive support for people that cannot trust automation or delegation, the task is too critical.

https://midombot.com/b1/home

I tend to draw the line at automating the LLM to respond to things. If it's responding to some sort of external source, that source is usually somewhat consistent to the point I'd rather have the LLM create a script to parse the data and do that automatically. I've got a job search tool that I built recently using Claude Code. CC created scripts to scrape certain websites and scheduled them using native OS schedulers. The results get parsed and dropped into a sqlite database. No LLM is involved in the automated portion of this process. I've got some general status scripts which push details about the current health state of my servers and apps and also will alert me when job listings reach some defined threshold. At that point I use the LLM to look through the new jobs and categorize them based on work I'd find interesting giving me a prioritized list.

If all LLM tools disappeared tomorrow, all of my scripts and processes developed with an LLM will continue to work without hiccup. If anthropic went out of business tomorrow, I'd lose nothing switching to another provider because I don't have to "trust" agentic operations in automated processes. They are always overseen by me and they are rarely creating things I couldn't have created myself. It's just much faster to iterate on it with these tools.

> If all LLM tools disappeared tomorrow, all of my scripts and processes developed with an LLM will continue to work without hiccup.

This is a really pragmatic philosophy and I think it's underappreciated. Using the LLM as a development accelerator rather than a runtime dependency gives the best of both worlds.

When you have a flow well defined, like transactions going on, it simply doesn't scale. But AI can then be used for analysis, alerts and investigating failures of such processes very nicely. Agents can also be used to prepare a transaction package that needs more human input, like a customer service case, but again with clearly defined outcomes. At least that's what I've seen in my limited experience consulting for a local online retailer.

That's exactly my process I follow now.

I look at the traces of agent execution, and use that as a feedback to extract common patterns. The comment patterns are extracted out as Scripts, or Skills.

So Agent doesnt have to figure out how to do things from scratch, saving considerable amount of tokens and latency.

I also came across this paper recently: https://arxiv.org/abs/2603.25158

Which does exactly the same. Extracts traces and converts them into skills for agents to use.

Everything is based on the requirements and available resources. One of our clients decided that calling the AI so often takes time, and money, and this does not work for him.

AI can give suggestions, not decisions. IF you want decisions and responsibility to be taken, use real people.