Aaron Lundsten

Faith, family, leadership, and AI. Air Force veteran, husband, dad, and founder of Proactive Strategy Coaching & Consulting. Helping businesses work smarter.

[Aaron Lundsten @LundstenA]

@sickdotdev No idea what you're talking about. Had a 5.6 ultra run go for over an hour and it only ate 2% of my usage...

[Aaron Lundsten @LundstenA]

@davis7 I think dropping them at the same time allowed them to manipulate the subagent routines that they needed to work on once everything dropped.

[Aaron Lundsten @LundstenA]

@mark_k I've been super happy with efficiencies so I'm not sure what's going on here.

[Aaron Lundsten @LundstenA]

@Dimillian Tibo announces he's going to bed and now everyone is posting their schedules on X?

[Aaron Lundsten @LundstenA]

The software most small businesses need won't take $1m to make. It may take a $200 sub and 1-2 months. The maintenance of that is still going to be lower than most subscription prices because software is no longer a one time fee.

Add an enterprise that pays millions per year for a software subscription...They find they can use $1m in tokens to build it over 12 months and it works exactly as they need, not how the software engineers designed it for the majority of it's users.

Software isn't dead but we are getting closer to a place where people will build what they need instead of paying someone to build it for them. Especially as agents become able to maintain. I use agents to maintain projects all the time. It often catches issues before people do and fixes them before anyone knows something was wrong.

Those who don't see this aren't using the tools that are at the front edge of what's available. Heck, I'm not using everything that's available so I found the one that works best for me and lets me build everything I need and maintain all. Instead of a subscription to 10 different softwares, I pay a subscription to one. And the best part, the one is not set so I can't go somewhere else if another tool becomes better.

I challenge people to go to Claude CoWork or my preference ChatGPT work. install it on your computer, install the chrome extension and turn on computer use. Give it a task like review my computer, email and anything else and let me know what I'm missing.

Most people will be surprised.

[Aaron Lundsten @LundstenA]

@Dimillian What work is on the docket then?

[Aaron Lundsten @LundstenA]

@cifilter I noticed that as well yesterday. They definitely cooked after the rough rollout

[Aaron Lundsten @LundstenA]

I typically have 6-10 pinned tasks. They are the tasks that I'm actively working on. Once the task is done, I unpin it. I have several daily agents that I keep pinned so I can review them as they do their heartbeats throughout the day.

My Chief of Staff thread keeps trying to get me to reduce the amount I have there so I may be merging some of my daily agents if it makes sense without context bloat.

[Aaron Lundsten @LundstenA]

I use the codex app constantly. I only use the CLI if codex crashes and all I tell the CLI is to get it back online(Windows is buggy).

I created a skill which I named Orchestrate.

In the skill it tells the thread that uses it that it's the orchestrator, I usually use Sol Medium or High.

Then I have a table that I had GPT put together that explains the strengths of the different models and reasoning levels based on Artificial Analysis reviews. Cost and intelligence are what it gave it.

Then it's instructed to use whichever model has the lowest cost for the level of intelligence the task needs and to break the tasks down into the simplest forms possible.

It runs for a long time and uses very low usage. I do spend some time before using it having it provide me a list of up to 5 recommendations to get the task done, up to 5 edge cases we need to consider, and up to 5 questions it needs me to answer before it feels confident. I'll do one or two of those loops until the scope feels right then it's off to the orchestration races to get it done.

Doing it that way, my tasks rarely take longer than 90 minutes once it start implementation to deployment of a feature after the scoping is done. shorter implementation time, less tokens used.

[Aaron Lundsten @LundstenA]

@PSwizard00 @thsottiaux @theo That's the theory based on Theo's original post. It would be nice, I'm hopeful with a sidebar change that may make a voice agent more able to control things.

[Aaron Lundsten @LundstenA]

I make sol the orchestrator. Outputs are the expensive part of agentic work. Sol reviews a lot then tells the other models to go do the work. The work is where most of the outputs happen. Then Sol orchestrator has explicit instructions to review everything and provide context to the agents if revisions are made.

It's like having a senior tech explaining things to someone that's new. The new person can stumble along, presents their work then the senior tech tells them what is wrong, how to fix it then tells them to go do it. That happens til Sol likes what was put together.

[Aaron Lundsten @LundstenA]

@Wicaodian @ForwardEditor The hardware and software are optimized but a couple of key workflows I use and apps don't exist...Now that I think about it, I could make the tools to do that...

Maybe I need to give it another shot.......

[Aaron Lundsten @LundstenA]

@thsottiaux I have a reset expiring in 3 days...I don't know that I can burn through this all fast enough.

[Aaron Lundsten @LundstenA]

I think their goal is more efficiency, more efficiency, more efficiency. I'll take a slightly less impressive model for a more efficient one any day.

Efficiency comes with lower costs and faster speeds, those make up for slightly higher intelligence even when you can use the models to their different specialties.

[Aaron Lundsten @LundstenA]

@CodeAkram @Dimillian Ai knew you were thinking about it, can't be lacking!

[Aaron Lundsten @LundstenA]

@ForwardEditor @Wicaodian I love iphone, ipad, apple watch, air pods...I cannot get behind mac. Every time I try...I gets so close to what I want it to be but falls short in the things that matter to me. I really want to love mac for the amazing hardware.

[Aaron Lundsten @LundstenA]

@TheRealAdamG I think it's extremely smart in the work and codex harness. I find it doesn't do quite as well in the chat harness. I get a lot more mistakes that I wouldn't usually get in work or codex. May be different models and reasoning since chat is light, medium, high, pro

[Aaron Lundsten @LundstenA]

An organizational harness. ChatGPT work is amazing but if I want to outfit my employees with it, there are some different levels that need to be handled. They could be with teams in different areas on the same page but we all know large corps aren't good at that.

[Aaron Lundsten @LundstenA]

@thsottiaux Next project if you're not already working on it. ChatGPT work harness that follows business guidelines and shares agents.

If I'm a corporation, Joe creates all these agents that do all this great work, then he leaves...I lose those agents.

A company harness where shared agents do work and when someone is out sick or transitions out, a framework that keeps those agents available to the business.