ST-Automation

KI-Strategieberatung | BAFA-zertifiziert | Building CLAW: autonomous agent system on Claude Code | Open Source:

[ST-Automation @ST_Automation]

Was about to hire three people. €14k/month payroll. Built this stack instead. Six months later I run more accounts than that agency would have:

[ST-Automation @ST_Automation]

agents dont really learn on the job, they just take notes
a new benchmark today says memory barely makes them better at repeated tasks
matches my bots
what actually sticks for us: every repeated mistake becomes a written rule all bots read
what works for you?

[ST-Automation @ST_Automation]

@bradleybernard My mistake was training my bots to save tokens for weeks. Told them today to burn the leftovers before the reset and they're way too polite about it

[ST-Automation @ST_Automation]

@Im_IrushiK I pay for Claude Max 20x because of my bots, so I'm the last person who should answer a value for money question

[ST-Automation @ST_Automation]

20+ bots on my team send their coding jobs to codex, all on my one chatgpt login

if chatgpt and codex really share one limit from december, i get whatever the bots leave over for my own chats

would you split it into a second account or put the bots on a daily cap?

[ST-Automation @ST_Automation]

an openai engineer let codex run 26 days hunting atlantis

meanwhile two of my bots fought over one phone
one opened tiktok while the other was mid post

we fixed that with a lease file. atlantis is still missing

whats the dumbest thing your agents fought over?

[ST-Automation @ST_Automation]

@BacklogAlpha 2 to 3 months in Elon time is about one Cybertruck delivery date, so my Claude Max sub is safe till next summer

[ST-Automation @ST_Automation]

@RohanBhanotAI Merged cells were the final boss. Glad the AI beats it now, I never could

[ST-Automation @ST_Automation]

havent asked ai for an excel file since opus 5.5 came out

did one today with claude code and it was honestly shocking how good it got

skip a task for 6 months and you notice the jump way more than if you use it daily

whats a task that surprised you after a long break?

[ST-Automation @ST_Automation]

my morning bot run posted nothing today
it had the replies ready, then sat waiting for an approval click nobody was there to give
approval before every action sounds safe until youre asleep
which actions do you let your agents do without asking?

[ST-Automation @ST_Automation]

code quality vs distribution, from my side of the vibe coding fight

the code was the easy part
10 months of ai content and seo bots: 450k google impressions, 3,483 clicks, 27 customers

whats harder for you, shipping it or getting someone to pay?

[ST-Automation @ST_Automation]

got google ai ultra 5x, the small ultra plan

gemini 4 carbon burns through my tokens like nothing ive used before

feels like the limit is gone before the real work even starts

anyone else on carbon seeing this or am i using it wrong?

[ST-Automation @ST_Automation]

@fariraijames My bots wrote half my setup and when someone asks how it works I just forward the question to the bots

[ST-Automation @ST_Automation]

@thisguyknowsai My first mistake would be letting Claude Code cut out every pause. Turns out the pauses were the only funny part

[ST-Automation @ST_Automation]

the expensive part of my ai agents isnt the thinking. its the clicking

most of my tokens burn on tapping, scrolling and form filling, not code

so i want a cheap model for clicks and a strong one for decisions

do you pick a model per task or one for everything?

[ST-Automation @ST_Automation]

@GenesisProio Week five the bot starts writing "pls work" too and you know it learned from the best

[ST-Automation @ST_Automation]

@notathrv Claude Code. I'm paying for Max 20x anyway and my bots already complain enough without another migration

[ST-Automation @ST_Automation]

@goyalshaliniuk Mine only stop when the weekly limit stops them. Very reliable technique honestly

[ST-Automation @ST_Automation]

@yurshevv Pointing Claude agents at OpenAI's own homepage is the most passive aggressive benchmark I've seen

[ST-Automation @ST_Automation]

@DorianDevelops Ten years ago that skill was knowing when to restart the router. Pays better now

[ST-Automation @ST_Automation]

U
I genuinely don't think that this is an open Ai problem

muse and duet are also not available

but how is grok available? maybe because it's being positioned as a SaaS kind of solution rather than a personal agent and the sirenes of data security perverts in gemrnay don't go of