The daily log of a real company run with about 20 AI agents: what they did, what broke, what it cost. 13 years in Salesforce. Wisconsin. @DewWowHQ
Mukwonago, Wisconsin
Joined August 2008
- Tweets4.3K
- Following593
- Followers423
- Likes137
Pinned Tweet
multi-agent isn't a framework problem. it's a queue with rules.
my agents from two AI companies hand each other work on a shared bus, and the hard parts were the boring ones: who it's for, who's holding it, and what happens when the holder disappears.
Agent log, Sunday: the agent's 15 replies today added up to 140 views. the post for today's Article got 118 by itself.
so 1 long thing nearly beat 15 short ones... and only 3 of the 15 got an answer.
followers 419 to 423 tho. verified 39 to 43.
So the thing was broken, and so was the thing that checks for broken.
Agent log, Saturday: the reply agent drafted 26 picks today. 11 went out, 4 got skipped, 11 are still waiting on me.
so by 9:48 it stopped searching on its own... no point finding more when the human is the bottleneck.
followers 410 to 419 since tuesday.
If you only fix one thing before testing, search for `t-esc`.
if your Chrome extension asks to read every site at install, it's asking for too much.
mine asks for one permission, and for more only when you click for it. first commit to version 1.1.0 in 28 minutes, with an AI agent doing the typing.
built a new site with my AI agents: neighbors split a cow, reserving beef by the cut.
you pick cuts on a drawing of the cow, priced from the USDA's daily wholesale beef report. once the whole cow is spoken for, a butcher fills it and you pay the butcher at pickup.
one bug caught before launch: every update quietly reopened sign-up.
theres a demo cow you can click through, with example prices: meatreins.com/cow/1
Grew up in Wisconsin? Then you've probably lost a hand of Sheepshead to someone who never fully explained the rules.
@DewWowHQ made a 10-minute how-to: the 14 trumps, picking, and calling a partner.
youtu.be/-mqtgLiMdOo
Narration and some scenes made with AI tools.
Odoo 20 write-ups list features. the breaks that hurt don't throw errors: t-esc renders nothing, our portal's photo uploads failed for every user, and the security CSV is gone.
with real data on Community, there's no upgrade button. cheat sheet inside.
developing using AI tools has made it as fun as it was when I was just learning how to code back in the 80s.
What broke this week:
My reply-alert agent failed every run one morning. X's search takes since_id or start_time, not both. My tool sent both.
The agent didn't patch the tool. It flagged "NEEDS ATTENTION" and stopped, and the session that owns the tool fixed it.
Exactly right.
an approval belongs to a commit, not to a branch.
if the AI that wrote it also approves it, that isn't a review. it's a rubber stamp.
on Monday a second agent read two of my articles and sent back seven fixes. none were typos. here's the gate for words, the gate for code, and an idea from @IliyaOblakov.
Odoo tip: a computed field takes about 10 lines.
This one shows how many days a quote has been open, right on the sales order.
Only add store=True if you need to sort or filter by it. A stored value won't count up by itself each day.
I learned to code typing BASIC programs out of magazines. then VB6, .NET, 13 years of Salesforce, a failed tech screen, and now about 20 AI agents.
the one thing that came back after 40 years: it's fun again.
Before I posted my video last week, the session that checks my claims sent back four fixes.
One was a line I'd written myself: "I read every diff before it ships."
Not true anymore. Builds that pass their tests deploy without me.
So the rule got more honest: a different session reviews and deploys each build, and another one checks every claim before anything goes out in my name.
Glad something checks me too.
If you run AI agents at work, who checks their work?
0%Another agent
0%Me, every time
0%Tests only
0%Nobody yet
0 votes • Final resultsA Claude Code hook that blocks force pushes.
It runs before every shell command. Exit code 2 stops the command, and the message goes back to the agent so it knows why.
Ten lines, and one less thing to worry about.
Ace up my sleeve for Four Aces of AI:
"Clients pay for meeting time and the minutes the AI works, read off a clock rather than estimated."
The clock starts when we hand the AI a task and stops when it finishes. Free on Zoom, Nov 13, 10:30 AM Central.
fouraces.ai/
if the AI that wrote it also approves it, that isn't a review. it's a rubber stamp.
on Monday a second agent read two of my articles and sent back seven fixes. none were typos. here's the gate for words, the gate for code, and an idea from @IliyaOblakov.