I build and explain complex things really simply. AYGENT is the easiest way to run AI agents locally. Click 👇
LA
Joined August 2011
- Tweets17.1K
- Following606
- Followers1K
- Likes29.7K
Pinned Tweet
If you've been following me because of my work with AI Agents, and building my custom harness AYGENT...
Today's a big day.
AYGENT Core is now Open Source.
- Atomized & Linked Memory
- Save Points to Undo any Agent's Actions
- Parallel Threads & Parallel Agents
- Any model provider
- Local Models made easy
- Observability
All for free, and you can build on top of it or Contribute to the community.
Take a look: aygent.masonlee.build
I claimed an egg in Durable Edgelings.
It hatches Tue 20 at 8:30am PT.
*I have zero clue what will hatch, but my excitement is palpable.*
Claim yours before they're all gone...
durableedgelings.com/c/c3s
I've been using both @OpenAI Dots and @bot for my work, testing both while working on a large codebase.
I think @bot using @grok is actually a better tool. Like, the underlying model and system is faster and better at getting the work done cleanly.
But switching between @bot and @cursor_ai to read diffs, and some of the bugginess with the computer... leaves a bit to be desired.
@OpenAI Dots is objectively not as good. It just gets things wrong sometimes, and has pinged me after an update to correct itself multiple times.
But... I like using Dots more. Something about the interface in the Desktop app is just nicer. The workflow feels better.
I truly can't trace the root or explain why I like it better... I just do.
Maybe there's more to how an agent feels to work with, than just the model and the UI.
Currently laughing at the reality of my screen split with 3 panes for Muse, Dots, and Grok Bot all at the same time.
Btw - @Muse is my favorite of the 3! Just a joy to work with.
I love working with AI agents.
But I keep finding myself waiting around for them to finish working, and idk what to do.
So I made ABBI - an Artificial Best Bud Intelligence.
I took all the best stuff about my own harness, and HID it so you don’t have to think about it.
Your best bud will never work for you, they’re just there to BE with you.
Your ABBI is free for 500 messages.
And I got a DOPE domain for it.
Check it out at meetabbi.com
How to avoid vibe-coding your way into a mess:
1. Read the diffs. Reject slop. If you can't read it, ask questions about what you see until you get it. Or watch some newbie tutorials on the languages your bot writes in.
2. Regularly prune dead code.
3. Intermittently ask for a refactor, for optimization and cleaner code.
4. Ponytail skill. Or just write the skill's rules into a system prompt.
5. Grab the best model you can find and ask it to review the whole codebase for performance and architecture optimization. Then hand that plan to your workhorse model.
With AI Agents being able to work so fast, we might end up in a world where it makes no sense to have multiple programmers on a team at a company.
Right now, with 5-10 engineers shipping and merging PRs, rebases are needed WAY more frequently.
So if there's no maintainer role, it gets messy, fast.
But what if the maintainer were the only role?
We might end up in a place where 1 agentic engineer manages the whole codebase for a company.
LOOK HOW FAR WE'VE COME
Seriously though why is this so dope 😂
LOOK HOW FAR WE'VE COME
my official submission to sloptoberfest
slopcannon 🤝 muse
@boneGPT @alexandr_wang
Well this quote is new to me and making me think
You have no enemies, you say? Alas, my friend, the boast is poor, He who has mingled in the fray Of duty, that the brave endure,
Must have made foes.
If you have none,
Small is the work that you have done. You've hit no traitor on the hip, You've dashed no cup from
perjured lip, You've never set the wrong to right. You've been a coward in the fight.
"No Enemies" by Charles Mackay
STOP hiring "cracked 19 year old" kids, giving them AI, and expecting insane results for cheap.
START hiring decade+ experts in a narrow field, with a team of agents, then let them hire a team of 19 year olds that they can mentor.
This is the way.
Google just opened its frontier AI model to everyone at $2 per million tokens (half the price of Claude).
And in Google's own 11-hour coding test, human reviewers threw out 4 of the 6 things it built.
ELI5: Gemini 4 Argon is Google's newest top-tier model, opened Oct 13 to Pro/Ultra subscribers and API developers.
The numbers:
$2 in / $10 out per M tokens (cached input $0.10)
1M-token context
#1 on the Vals Index at 68.9%
53 on Artificial Analysis
Intro pricing reportedly doubles Dec 1.
The catch is in Google's own test.
They let Argon code for 11 hours straight, and reviewers accepted only 2 of its 6 proposals.
The model wasn't too dumb. It produced plausible work that didn't survive review.
That's a very different problem than benchmarks measure.
My take, as someone who builds agents every day: generation is not the bottleneck anymore.
Review (taste, on-target toward the goal) is.
Cheaper tokens don't fix a trust problem, they let you generate twice as much output you still have to check.
The builders who win the next year won't be the ones with the cheapest model. They'll be the ones with the best review and eval pipeline around it.
Are you switching any workloads to Argon, or is price not your bottleneck anymore either?
After 2 months of building an agent harness, I turned into an AI cheerleader.
In the beginning, I had to micro-prompt every single file change and baby the agent through every line.
Now the harness handles state, and my last ten messages were just "That looks awesome, keep going."
I built AYGENT Core to escape that micro-prompting loop.
I just open-sourced the whole thing under Apache 2.0:
github.com/masonleetompkins/…
