Why does anyone read social media anymore? It's certainly not to learn anything... Yet in my case that's mostly true, and I think I'm in the same boat as many others.
We use social media as a place to find answers to very particular questions, usually relating to experience or what the rationalist people call "tacit knowledge." For most of my life I like to think this has been the case, although I have to admit I've recently developed a bit of a Twitter addiction. Not badly—I only ever use it in a browser. I have maybe 300 posts in my lifetime, but I do like to read others' posts. That short-form content has got me.
In the more annoying corners of what I suppose I ought to call "𝕏", they call the thing I seek "alpha," as in Seeking Alpha™... or essentially an "edge", if you aren't familiar with trading parlance.
No matter what I tell myself about the health of social media in any form, nothing quite answers a question like X. I read replies too, sometimes. Unfortunately (or fortunately), I've discovered many good ideas from that site, but it has been irritating me so severely over the last few months. Mind you, I never really used it before about 11 months ago.
How Shell I Replace You?
A bot. My bot. Ladies and gentlemen, I introduce to you today: 24-7-B0T.
This bot, or B0 as I'll now call them, is my first fully custom, in-house-orchestrated agent stack. Technically "they" are a team, but I like to name things, so it's only logical to call them B0 for "bot zero"... The T is for Twitter (because "X" is still a dumb name).
B0 solves my conundrum
Who could resist the very real alpha we get from instant updates on the things we care about? Obviously some people choose enlightenment (and avoid Twitter) like I once did, but I want that data.
The problem is it's still social media, so realistically, no matter who you are, your brain is getting hijacked bro. I know that mine is.
It should be noted that this post is the first one to be drafted with pen and paper. Could you tell?
Bright idea: bring the posts to you! Screw visiting the website. No more data for the algo to better target me, locking me deeper, and I get to keep my sweet edge! This would be simple enough if following people worked. But it can't because:
a) The people you follow aren't posting worthwhile stuff all the time.
b) The "For You" tab kind of slaps, actually.
So what better way to do sentiment analysis and extract worthwhile posts than crawling your timeline with an LLM? Yay!
Here it comes
BIG OL DIAGRAM HERE TBD
The stack consists of a couple of tools I recently found. The primary harness is AIO Sandbox from ByteDance which is running inside of an Apple Container. The means of clicking around the timeline and navigating the site is Google's Gemini 2.5 CUA, just to try it. It's kind of pricy, but wow is it amazing to not have to scroll, hunt for selectors, and worry about Nikita changing those or any anti-bot measures. It's too expensive for anything that would run more often than a few times per day though... The agent piloting both is Gemini Flash; I'm also planning to try driving with the DeepSeek team's stuff because I've got a bunch of credits kicking around.
Pilot initiates a request to load "For You" and an initial screenshot is taken. From here we start our agent loop. According to my personal DB of bangers, the visible posts are evaluated. If a good candidate is found, two things happen: 1) the post's source link is logged along with any attachments and content; 2) the Pilot dives into the thread (in another tab).
After the rest of the thread (really the top 10 replies and the exchange with the OP) is read, the complete summary-construction routine is called. This deposits the thread parts into a safe place should they be needed later—people do delete stuff often. After constructing the thread backup, we go back to the main tab by closing this thread tab.
The Pilot continues hunting, only clicking into likely bangers or high-alpha posts. He scrolls. He loops. He scrolls for a while more—about 10 minutes—then reloads "For You." The Pilot also inspects the "Following" tab, but it's much more cut-throat: a higher threshold is required before clicking into anything that might be interesting.
After a session, our results are coalesced and ranked. In that order, they're sent as a blob to my endpoint for agent orchestration—command and control, whatever.
Now I can check my own website instead of Twitter. The posts are "published" so they can be embedded for proper viewing if they contain video, which saves bandwidth and disk space.
"For You" as an API would be nice, but I'm already dodging their ridiculous fees for API usage, so we'll see how long this lasts. I wouldn't mind getting a ban anyway.
🛫 Away on business
A selection from this method will be available at
/agents... Soon.
The agentic approach
What I'm working on next:
- Trading agent - to remove human hesitation from profit-taking decisions
- Backtesting agent - to test ideas against history instead of my memory
Something tells me these residential web-browsing agents are going to be useful there.