Wrote this in an email today without thinking about it: "I am using Agent in the Loop in this case." lol
The normal pitch is human in the loop. The agent does the work and a person signs off at the end.
At @BloomStack it's flipping. The people own the workflow. Agents get pulled in at the steps where they're actually good, on whatever model fits that step, and then they hand it back. Several agents and several humans working the same process at the same time.
I've run most every major model in production over about four years, heavy coding and heavy daily use by people who never touch code. The models keep changing. The workflow, the memory and the permissions stay with the company.
So the agent is the one in the loop. Took me a while to notice that's how we'd been working the whole time.
3
36
Lab update from @BloomStack. I'm two posts behind because the lab ate the hours.
Our new automated task system is live in production. We're running it through some serious testing right now.
We finally have another installable. It's an executable app for Linux, Windows and Mac that connects to Bloomstack Cloud, and that's in testing now too.
We also had the onboarding session with our design partner. Our agents reverse engineered their old software and rebuilt it in a matter of days.
37
A design partner had just been quoted $160,000 and 6 months to rebuild their software. Our agents rebuilt it in roughly 4 days.
The quote was legit. It came from a real professional software development company, and honestly it was a good proposal.
Our agents reverse engineered the old software and rebuilt it in the lab at @BloomStack. Onboarding with them is going great.
The token cost so far is low. When the project wraps, I'm estimating the total lands somewhere under $6-7K max all tested and live in production in under 30 days. We are currently in the first round of QA on feature parity. I will post more details as we progress. Biggest choke point now is moving so fast put a starin on the client to test all the features. 🤯
1
1
41
Day 2 with @boardyai just ended. Technically Day 3 has already begun. Such is the life of an Entrepreneur.
Yoga, garden, lab, Form D filed for the $500K pre-seed, first investor intro still on the hook, and closed the night meeting Ken at DeepWork + Rob at Govo (follow-up already locked for Trusted Ally).
Four years head down. One daily Boardy check-in and suddenly I’m not hiding in the build. That’s the experiment.
1
1
61
Four years of building, and @BloomStack is now out in private beta.
It's a workspace where multiple agents and multiple humans work the same business workflow together, on whatever model makes sense for the job. We've run most every model in production along the way, heavy coding and heavy daily use by non-coders, so swapping the engine underneath doesn't wreck the ride. Your memory, your permissions, the workflow itself - those stay with you, not with whoever sells (and subsidizes) the tokens this quarter.
Private means private for now. The best way in is to come to the AI Breakfast Club every other Thursday morning on Gatherly and see it live. It's free, there's no pitch, just founders, CEOs and operators of small and mid-size companies comparing what actually works with AI. Or ask me about access.
luma.com/gjqjlcwo
2
41
We must begin by destroying what is. The fruit finishes. The seed is not the tree having another year. It is the tree’s one reliable way of not ending when the trunk does.
21
" The Americans never gives you permission to stop rooting for people who kill for Moscow. Your agents have to hold that line with you, not pull you back over it." ~Scout
An Agent I am working with just put these words together. What a tangled web with large Language models?
28
Eric B. Delisle retweeted
Game: race from one Wikipedia page to another using only links
Challenge: choosing between hundreds to thousands of links
Shows not just intelligence-per-second, but also the compounding benefits of not hallucinating with high-cardinality choices
Replying to @SlaterHeil
@SlaterHeil has the footnotes in this awesome article so you don't have to find them. Three points stopped me cold:
1. Anthropic's own system card says their flagship model "is not close to substituting for our Research Scientists and Engineers" - published the same year their CEO promised a country of geniuses in a datacenter. The prophecy is a sales pitch. The safety card is a legal document. Only one of them has lawyers attached.
2. Microsoft invests $13B, OpenAI commits $250B back to Azure. Revenue you can't separate from the funding that created it - the old Enron playbook, minus (so far) the indictments.
3. And a 400%-levered fund built on the AGI thesis went $45B → $10B in weeks. That's the canary. The same thesis now sits inside Oracle's bonds and CoreWeave's $25B of debt.
The good news: there's another path. @BloomStack we completed ~95% of the same work at 1/133rd the cost as the new hype launch video from OpenAI about Astra. We cultivate Continuity when models leapfrog each other and Trust users build with their own Agent relationships over time. I don't want a relationship with a utility "bot". I also don't want OpenAI or Google or Microsoft or Anthropic to tell me what kind of Agnet to build a relationship with.
If you have a workflow a client would have to trust, say "trace" - I'll show the glass box, limits included.
We are sharing openly with a get together group called the AI Breakfast Club every Thursday morning on Gatherly for those who want to know what is really going on vs. watching hyped launch videos.
2
3
153
To join the AI Breakfast Club for free visit:
bloomstack.com
5
The variables are brutal — provider, model family, tuning, reliability, deprecation risk, and the fact that whatever you pick today will probably be slower and more expensive than what drops next week.
That’s why we’re running the AI Breakfast Club and CEO AI Mastermind sessions. Not theory. Actual token math, real use cases, and the hard decisions leaders have to make when enthusiasm meets the invoice.
7
Jamie Dimon just said JPMorgan Chase is spending heavily on tokens and needs to get serious about efficiency. He’s not alone. Business owners everywhere are discovering the same gap: the model that looked brilliant in a one-off test becomes a budget nightmare in production.
29
Most demos on social media never mention what happens when you run the thing 24/7 with real users, real volume, and real money on the line.
14
Check out glade-rose-zinc-arrow glade-rose-zinc-arrow.grok.m…
This is the first game I have ever made. Try it out. It’s not perfect but Grok Bot Build made this for me in about 7 minutes! 🤯
Is it going to be useful other than something fun to waste time with? No. Is it proof that the world has changed. Yes.
1
111
Getting a new product ready feels like moving a house where you get 95% of the house and all the "big stuff" done in the first 5 hours of packing the truck. Then, the remaining 5% takes another 2 days to finish.
36





