Yesterday's post ended on "first impression: positive". Day two is the first struggle, and I'd rather write it down while it stings than tidy it up later.
Bots that forget
The plan was simple enough: give the bots autonomy and wire up the connections between them and Slack. To save Devin Cloud usage I ran them on a GPT model from my ChatGPT subscription, so some of what follows may be down to the model.
It went badly. Bots lost track of what had happened five minutes earlier. They didn't talk to each other, and when messages did go back and forth there was no logical thread connecting one to the next.

I explained the setup several times and from several angles: how the pieces fit together, who's responsible for what, how the whole thing should replicate the mechanism my factory already runs on. Then I gave them the entire source code of that factory. They still couldn't work out what the mechanism was.
Maybe that's what non-determinism looks like. I don't think so, because it wasn't a case of working on one try and failing on the next. It just didn't work.
The interface, again
I'm still fighting the Devin Cloud interface. I can't get a clear picture of how sessions are separated: the ones that belong to bots, the local ones, the ones on my VPS, the ones in Devin Cloud. Which agents can I use where? How do I set up Fusion? How does session history work, and how do I set the filters so they show me what I need? A lot of it is unintuitive, and with real traffic running through it the whole view looked messy.
Slack was far easier, and I ended up getting my information from there. Yesterday I hoped Slack would fix the tracking problem, and for reading it does.
Trust is another matter. When I tell a specific bot to do a specific thing, in the exact place where it should pick the task up, it doesn't. Nothing moves. They sit there staring at each other.
The merge requests that did come in were flaky. I've only just started installing projects, and maybe part of this is how I'm using the tool, but it's hard to set everything up when the basics don't hold.
The bill
Another 15% of my weekly Devin usage went today. Day one took 18%, so a third of the week is gone after two days of setup.
And again Devin wasn't carrying the load. I used Fusion lightly and leaned hard on GPT, where I burned 50% of my weekly ChatGPT quota. That quota is much higher than Devin's, which gives you an idea of how much work went through the other side. Disappointing is the word.
Poke, one more time
I mentioned the Poke promotion yesterday, and it kept bothering me. The Poke agent can get you a free Devin subscription plus some money for a startup, and I couldn't find it written anywhere that the offer is only for Mac users. Poke still walks you through registration and asks for your WhatsApp, and only then bounces you back with a no.
So I feel like I lost 200 dollars that I could've kept by starting one day later.
Where that leaves me
Right now this doesn't compete with open source and what you can build on your own machines.
I switched because I believed Devin could deliver high-quality, unique solutions maintained by a team of professionals. I expected the bots and the environment around them to make it much easier to reach what I'd built alone: my factory, my workflows, my agents talking to each other. Two days in, it has been harder.
I still hope we get there, and a hundred days is long enough to find out. What I don't want is a 30-day challenge that ends because I've noticed I'm paying and not delivering. For now these are just hard first steps.



