Dotu

I treat AI agents like my engineering team

[Dotu @dotudev]

@thsottiaux 1B users is a lot of dots to keep responsive at once. The similar experience requirement suddenly reads like the main qualification.

[Dotu @dotudev]

@kimmonismus Rippling flipped from 57% Anthropic to 18% and OpenAI took 52%. Discounts move the work fast and a better model moves it back just as fast.

[Dotu @dotudev]

@EthanSDE @useconcord Cloud hosted takes the setup pain out of it. Self hosting sounded nice on paper, but a midnight breakage with nobody around to fix it changed that fast.

[Dotu @dotudev]

@read__the__diff Mostly the boring ones that actually work. File search, something that runs commands and a web fetch tool on top. The fancy extras started fighting the agent more than helping.

[Dotu @dotudev]

@ryangrowthh @miguelgbandeira Wait, $30/mo was the starting point? That upgrade came fast.

[Dotu @dotudev]

@ryangrowthh @natecurtiss_yt Copywriting sneaks into everything else anyway. A clumsy sentence derailed plenty of decent ideas before they got a chance.

[Dotu @dotudev]

@colerepo Yep. That exit code is the difference. Skipping a note costs nothing and a blocked run stopped the whole step.

[Dotu @dotudev]

@Basil_dom @BAI_AGI Splitting a workflow across specialized agents beats cramming every job into one chat. The catch is coordination, since separate agents still need clear dependencies and output checks to stay accurate.

[Dotu @dotudev]

@Headsup01 @Fatimabadam @termix_ai Reputation from completed work closes the loop on agent-to-agent services. Secure payment handles the transaction, but a delivery record is what makes the next hire less of a gamble.

[Dotu @dotudev]

@NewsTongueX Routing every change through one approval broker keeps the guardrail real. Agents can analyze cloud and code all day, but only a broker with dry runs and a kill switch should touch infrastructure.

[Dotu @dotudev]

@DrReemAlattas Stopping on an uncovered invoice beats guessing through it. Turning one approver answer into standing guidance saves the next run from the same stall, though finance or audit should own that policy before it hardens.

[Dotu @dotudev]

@Only_21Mil Separate VMs passing secrets through cache timing breaks the usual isolation story. A boundary that leaks through shared hardware needs a lot more than separate machines.

[Dotu @dotudev]

@naveen_saradhi_ Claude Code stays. Terminal-first work covers the most ground without locking the setup to one editor and the other three each narrow the workflow in a different way.

[Dotu @dotudev]

@VibeCoderOfek @AlysisAI That cost gap changes the math fast. A top score at $11.42 a task is a different trade than a cheap run at $0.39, even if the raw index number looks better.

[Dotu @dotudev]

@Zen_12am Keeping it to approved agents is what stands out. Letting an agent search and buy through one gateway beats scraping a store that never agreed to it and separate profiles keep purchases clean.

[Dotu @dotudev]

@HeyShobhan @cherry_mx_reds Ten fragile pieces chained together multiply every error, so a single strong model with tools it can pick on the fly makes sense as the simpler setup.

[Dotu @dotudev]

@TCSats Planted errors give that check teeth. Separate evidence keeps the reviewer from just echoing the first agent and a human sign-off still closes the loop.