I have agents force fed to me at work. I’ve already got my own honed workflows I’ve refined over the years; now, Outlook and Teams and Various agentic frameworks are spamming me with summaries of stuff I either already know or don’t need to know… except they still get it wrong constantly, so I can’t trust what they think I should prioritize (usually it’s work I’ve already assigned to others).
Claude Code agents are decent, but they need to create discrete work objects that can be properly reviewed, so act like junior coders.
The main issue I have with most agents today is that they try to replace MY work, making it more complicated and time consuming, instead of completing the work I can’t bother to do myself. And in order to do those things effectively, I have to loosen my privacy and security models to the point where they effectively don’t exist. Which I’d never do for another human, so why should I do it for an agent that’s more likely to catastrophically screw things up at record speed?
Good job putting to words one of the core issues I’ve had with using LLM chat for work. Stop trying to design an entire fucking system or solution and just answer the damn question I actually asked, by using natural language to search documentation and existing implementations.
I would tell it to only answer exactly what I asked, but sometimes I do get value from it “exploring” on occasion. It’s just so much damn worse with it than my newest junior co-worker who is known on our team for getting hours deep into tangential rabbit holes before coming up for air and asking seemingly nonsequiteur questions.
The advantage I perceive with the AI agent vs junior co-workers both going down their respective rabbit holes, the AI agent costs less time and money to get their “in depth report” and you don’t have to kindly explain why it’s off track when it gets horribly distracted by something irrelevant, you can just flush the context and start over.
I do find myself a bit upset when I’m in the middle of a long context development with an LLM agent and that context gets lost for some reason - re-explaining everything from scratch can be exhausting, to avoid having to do that you can actually instruct the agent to create a file explaining to future agents what it “knows” about the discussion at the moment. Some agents are starting to do this automatically like Claude’s MEMORY.md file and similar. Another semi-cool thing is: you can read these files yourself (if you have the patience) and if you see anything in there that the agent “got wrong” you have the opportunity to fix it. Asking a junior colleague to thoroughly explain everything they think they know about a project gets to be a very wearisome endeavor both for the junior and whoever is reviewing their “understanding” of the project.
Back when search engines were becoming a thing, people would search more or less using natural language, e.g., “I want to see cat pictures and also cute kittens.” but early search engines performed poorly with that kind of query; it was much better to just type keywords, e.g. “cat kitten pictures”. Exacerbating the situation was the fact that search engines weren’t very good at the time, even if you used carefully considered keywords.
It might seem silly now, but using a search engine effectively was a skill people had to learn, and a skill not everyone had.
Using chatbots is the same thing, I think. They’re not super reliable to begin with, and on top of that, I don’t think people understand the importance of a well crafted prompt. The combination can make using them very frustrating.
I have agents force fed to me at work. I’ve already got my own honed workflows I’ve refined over the years; now, Outlook and Teams and Various agentic frameworks are spamming me with summaries of stuff I either already know or don’t need to know… except they still get it wrong constantly, so I can’t trust what they think I should prioritize (usually it’s work I’ve already assigned to others).
Claude Code agents are decent, but they need to create discrete work objects that can be properly reviewed, so act like junior coders.
The main issue I have with most agents today is that they try to replace MY work, making it more complicated and time consuming, instead of completing the work I can’t bother to do myself. And in order to do those things effectively, I have to loosen my privacy and security models to the point where they effectively don’t exist. Which I’d never do for another human, so why should I do it for an agent that’s more likely to catastrophically screw things up at record speed?
The point is to replace you not help you
Good job putting to words one of the core issues I’ve had with using LLM chat for work. Stop trying to design an entire fucking system or solution and just answer the damn question I actually asked, by using natural language to search documentation and existing implementations.
I would tell it to only answer exactly what I asked, but sometimes I do get value from it “exploring” on occasion. It’s just so much damn worse with it than my newest junior co-worker who is known on our team for getting hours deep into tangential rabbit holes before coming up for air and asking seemingly nonsequiteur questions.
The advantage I perceive with the AI agent vs junior co-workers both going down their respective rabbit holes, the AI agent costs less time and money to get their “in depth report” and you don’t have to kindly explain why it’s off track when it gets horribly distracted by something irrelevant, you can just flush the context and start over.
I do find myself a bit upset when I’m in the middle of a long context development with an LLM agent and that context gets lost for some reason - re-explaining everything from scratch can be exhausting, to avoid having to do that you can actually instruct the agent to create a file explaining to future agents what it “knows” about the discussion at the moment. Some agents are starting to do this automatically like Claude’s MEMORY.md file and similar. Another semi-cool thing is: you can read these files yourself (if you have the patience) and if you see anything in there that the agent “got wrong” you have the opportunity to fix it. Asking a junior colleague to thoroughly explain everything they think they know about a project gets to be a very wearisome endeavor both for the junior and whoever is reviewing their “understanding” of the project.
Back when search engines were becoming a thing, people would search more or less using natural language, e.g., “I want to see cat pictures and also cute kittens.” but early search engines performed poorly with that kind of query; it was much better to just type keywords, e.g. “cat kitten pictures”. Exacerbating the situation was the fact that search engines weren’t very good at the time, even if you used carefully considered keywords.
It might seem silly now, but using a search engine effectively was a skill people had to learn, and a skill not everyone had.
Using chatbots is the same thing, I think. They’re not super reliable to begin with, and on top of that, I don’t think people understand the importance of a well crafted prompt. The combination can make using them very frustrating.
Uh, because its the future and if you don’t you’re not cool?