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.
> 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.
More I use AI tools, stronger I'm convinced that it's a force multiplier. I'm one of the strong advocates for adoption of AI at work.
But I'm also very skeptical about the narrative- that AI will simply replace workers.
The main issue is accountability. If an autonomous agent takes an incorrect action, who takes responsibility?
I recently had a first hand experience at work where an agent, designed to act on customer tickets, was authorized to suspend accounts upon request.
It incorrectly suspended an active, critical account essential to our revenue metrics. Now, the support engineer who deployed that agent is writing the postmortem/CoE.
These are some incidents why I believe AI will not "completely" replace human roles. When systems fail at scale, we still require an accountable human to analyze the failure, accept responsibility.
I have a lot of questions after reading this article
1. How would Cocaine end up in wastewater treatment plant in the first place?
2. What is the volume of Cocaine needed, or used to observe this behavior from the fish?
3. If Cocaine pollution has already been happening in the wild, are there any recorded events that support the claim of effects on the habitat
> In the end, the “shortcut” had cost her more time than if she’d just done the work herself from the start.
I had this same opinion on AI IDE Copilots about a year and a half ago. They were too nascent, and writing code manually saved me hours of debugging their buggy code recommendations.
Fast forward - today—IDE Copilots have grown leaps and bounds in its quality of outputs. They have real utility now.
It's important to note that - "This is the worst AI agents will ever be, it will only get better moving forward."
I'm confident these tools will keep improving and eventually create net productivity gains, including for the Excel use case you mentioned.
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.