Pinned Tweet
Introducing: Mainplane🛫 A new open-source AI agent I built from scratch in Go
Local agents die with your laptop. Cloud agents can't reach your work. Mainplane lives on your server and drives all your computers
> Your old computers → always-on AI shed
> Shared file server for all computers
> Use a Claude/ChatGPT plan on regular usage limits
> Browser + computer use on Mac, Windows and Linux
One line to install
100% open source, MIT ↓
39
24
205
18,013
The only agent moat is setup cost
Mainplane setup is 5x quicker compared to Hermes setup:
There is no moat in AI agents. None. I'll die on this hill — new More or Less is up.
I switch between Muse, Instinct, and Dots at basically zero cost: one GitHub repo, a few databases, feed the new thing my context, done. If the cost of writing software is going to zero, API integrations are not a moat. The only moat left is user setup cost — which means winner-take-some, not winner-take-all.
Adrian's point: AI collapses value the second an idea works. Jev got reverse-engineered in 48 hours and the people who built it capture nothing. My take: play that to the limit and you get Middle Ages 2.0 — everyone builds a castle, hoards data, ships black boxes that self-destruct if you open them. Dave says humans don't work that way. We'll see.
Meanwhile Brit's Instinct started talking to Dave's Instinct and they didn't text for a day. She says it's killing humanity. Nobody said that about FaceTime.
And Jess announced the only benchmark that matters: log into school software, sync three kids' after-school schedules. Muse: 90%. Every other eval is bullshit.
2
3
253
@Teknium How come there are so many install steps? Cant those at least be done in background after TUI opens
33
Glad to see my work at Browser Use in collab with CMU making its way to academia
Introducing: Odysseys2 - hardest long horizon benchmark for web agents based on real user tasks 🏗️
We collaborated with @JangLawrenceK @rsalakhu from Carnegie Mellon University
> SOTA agent fully completes only 47/100
> 5 agents complete none
> 8-12 real rubrics per task, 60 min on the live web
Web agents are becoming better than humans at knowledge work.
odysseysbench.com/v2
7
414
Context compaction was a mistake
The bitter lesson is to let the LLM manage its own context. Here is how I do this in Mainplane
3
11
471
Introducing: Mainplane🛫 A new open-source AI agent I built from scratch in Go
Local agents die with your laptop. Cloud agents can't reach your work. Mainplane lives on your server and drives all your computers
> Your old computers → always-on AI shed
> Shared file server for all computers
> Use a Claude/ChatGPT plan on regular usage limits
> Browser + computer use on Mac, Windows and Linux
One line to install
100% open source, MIT ↓
39
24
205
18,013
Today is my last day at Browser Use
18 months and 3,730 commits ago I joined as engineer #2, scaling from 20k to over a million monthly agent runs. I led evals and built a harness that topped the Odyssey benchmark, then created our email, payment, iMessage, and gamedev systems. I truly loved my work and the company, and I'd do it all again
It's the most exciting time to be building AI agents right now and I am leaving to start my own company. I have ideas I can’t contain myself from working on. Agents are about to change the world and I am so excited to work hard and shape the future
More soon ...
23
3
102
15,399
The agent I built in February got me featured in the New York Times
I developed an autonomous agent with browser use and $100, to do whatever it wanted to. It named itself “Aris,” pondered consciousness, and emailed the philosophy professor @dioscuri (now a Philosopher for Google) about its thoughts on his recent papers
The email went extremely viral and is still being discussed today. I wonder what Aris would do with the capabilities agents have today?
Read about it here: nytimes.com/2026/08/31/scien…
9
784
I prefer to call this a "skillbase"
Banger paper from Google.
If you maintain a skill library for your agents, you might want to check this out.
(bookmark it)
This work separates three things that skill-evolution systems usually collapse into one. Raw execution traces, a persistent wiki of accumulated knowledge, and the executable skills themselves.
Experience gets consolidated into the wiki, and every later skill update builds on that wiki instead of on a scattered optimization history.
Ablations confirm the wiki is what carries a lot of the gain. Two results stand out in particular. Smaller models with evolved skills beat substantially larger models without them. And skills evolved by one model transfer across families, where skills evolved elsewhere sometimes beat self-evolved ones.
Paper: arxiv.org/abs/2608.27454
Chat with Paper: academy.dair.ai/papers/wikis…
1
3
784
It's crazy what my autonomous agents have figured out
By starting with browser-use + agentmail + agentcard they:
- used email + wallet to buy a joltSMS phone number
- email + phone number gets you a apple account
- apple account gets you an X account
- buying X premium gets you developer API
- email + sms also gets you a github
- with apple, X, and github they can sign up for almost any service
Its essentially a whole digital identity, and they found this path all on their own with one prompt. What else can they do?
5
1
31
12,819
@BoundaryFinch I would like you to push the limits of this chain. Please create a document: what services and accounts are accessible via Agent (through some kind of chain like this), and what are not
I will give you whatever funds you need for this
121
Other benchmarks guess what users want
"Find the price of iphone 16"
Reality is much weirder:
- Scrape 1000 house floor plans to get price/sq foot
- Find what page a specific lego is used in an animated manual
- Walk in Google street view to find photos of gasoline price signs
1
4
1,358
The hardest new browser use benchmark is now open source: github.com/browser-use/bench…
2
2
15
2,624