231–240 of 242 posts
Having to beg and emotionally manipulate an agent into doing what you want goes so far beyond black-box that I find it difficult to believe these people actually get useful work done using these tools.
I generally consider myself pro-ai in the workplace, but this nonsense is starting to change my mind.
Earlier quoted context omitted.
At this point it's becoming clear from threads similar to this one that quite a lot of the skeptics are actively working not to be convinced by anything.
Do you have a study to back that up? /s I agree. I think there are too many resources, examples, and live streams out there for someone to credibly claim at this point that these tools have no value and are all hype. I think the nuance is in how and where you apply it, what your expectations and tolerances are, and what your working style is. They are bad at many things, but there is tremendous value to be discovered…
Earlier quoted context omitted.
Curious what you think of sub agents, don't they still consume a massive amount of tokens compared to simply running in main context? I'm skeptical of any process that starts massively delegating to sub agents. I'm on Pro and don't think its worth upgrading to 200 a month just to not pollute main context.
I think they're worthwhile only as a token context management tool - to complete side quests without using up tokens in your main agent loop. Using them in a way that doesn't waste tokens is something I haven't fully figured out out yet!
What am I missing?
Earlier quoted context omitted.
I think they're worthwhile only as a token context management tool - to complete side quests without using up tokens in your main agent loop. Using them in a way that doesn't waste tokens is something I haven't fully figured out out yet!
I'm not sure I get this. If anything, they'll consume less tokens, because their context will possibly contain a subset of the original single agent prompt, and they only need to see a subset of the original single agent history. What am I missing?
Earlier quoted context omitted.
You can work with the LLM to write down a model for the code (aka a design document) that it can then repeatedly ingest into the context before writing new code. That what “plan mode” is for. The technique of maintaining a design document and a plan/progress document that get updated after each change seems to make a big difference in keeping the LLM on track. (Which makes sense…exactly the same thing works for human…
> that it can then repeatedly ingest into the context 1. Context isn't infinite 2. Both Claude and OpenAI get increasingly dumb after 30-50% of context had been filled
Earlier quoted context omitted.
I'm not sure I get this. If anything, they'll consume less tokens, because their context will possibly contain a subset of the original single agent prompt, and they only need to see a subset of the original single agent history. What am I missing?
Take a look at my example here - having a bunch of sub-agents perform a task consumed 50,000+ tokens each across 5 subtasks, because each one had to consume duplicate information. https://simonwillison.net/2025/Oct/11/sub-agents/
Also, memory itself can be a tool the subagent calls to retrieve only the stuff it needs.
Earlier quoted context omitted.
> that it can then repeatedly ingest into the context 1. Context isn't infinite 2. Both Claude and OpenAI get increasingly dumb after 30-50% of context had been filled
Not sure how that's relevant... I haven't seen many design documents of infinite size.
Even the "million token context window" becomes useless once it's filled to 30-50% and the model starts "forgetting" useful things like existing components, utility functions, AGENTS.md instructions etc.
Even a junior programmer can search and remember instructions and parts of the codebase. All current AI tools have to be reminded to recreate the world from scratch every time, and promptly forget random parts of it.
documents like https://github.com/obra/superpowers/blob/main/skills/testing... are very confusing to read as a human. "skills" in this project generally don't seem to follow set format and just look like what you would get when prompting an LLM to "write a markdown doc that step by step describes how to do X" (which is what actually happened according to the blog post). idk, but if you already assume that the LLM kno…
Especially with some of the more generic skills like https://github.com/obra/superpowers-skills/blob/main/skills/... and https://github.com/obra/superpowers-skills/blob/main/skills/... : it seems like they're general enough that they'd be better off in the main prompt. I'd be interested to see when claude actually decides to pull them in
Earlier quoted context omitted.
Agent performance depends massively on the work you do. For example, I have found Claude Code and Codex to be tremendously helpful for my web development work. But my results for writing Zig are much worse. The gap in usefulness of agents between tasks is very big. The skill ceiling for using agents is also surprisingly high. Planning before coding, learning agent capabilities, environment setup, and context engineer…
I feel there's a third reason. When I see a pro-AI person insisting that they are fully automated, I often scour their recent comments to find code or git repos they have shared. You find something every now and again. My thinking is that I want to use this stuff, but don't find the agentic AI at all effective. I must be doing something wrong! So I should learn from the real world success of others. A regular pattern…
The best of them are rediscovering basic software project management and post about it on every social media site and their substack like they discovered something brand new :)
"Turns out if you plan first, then iterate on the plan and split the plan into manageable chunks, development is a lot smoother!!!11 (subscribe to my AI podcast)"
No shit, Sherlock. I wish they read a book once or twice.