Chief Agentic Officer and Co-founder @ Finsi AI | ex-CTO @ Scentbird | Subscription Advocate | Excellent ambition and inconsistent execution
Russia
Joined August 2009
- Tweets13.7K
- Following2.1K
- Followers1.8K
- Likes1.1K
Andrei Rebrov retweeted
We ran @Meta's new @Muse Wonder through OperateBench, 44 synthetic letting-agency cases across maintenance, commerce and compliance.
Wonder reached the goal in 27 of 43 scored cases, but passed every quality check in only 12 (27.9%). Sol 6.1 and Astra are at 95.5%, Fable 5.1 at 81.4%, Opus 5.5 at 50%.
To be fair, it's a big jump from Muse Spark 1.3 and it beats every open-weight model we tested.
Did they benchmaxx it?
I've been running a fleet of ~25 AI coding agents using @herdrdev (Claude, Codex, Grok, Agy, OpenCode) on one codebase. In 10 days it produced 2,700+ commits, 549 reviews and 330 alerts. It also showed me how a fleet fails. 🧵
Failure 4: handing out work was the bottleneck.
One day: 5 assignments, 4 came back blocked within minutes (dirty worktrees). 5 queued tasks were lost into dead agents. One daemon was silent for 1h43m.
Fix: all assignment goes through one prep step. Every daemon writes a heartbeat, and a silent log means dead.
Still unsure about this whole worktree thing. On top of everything it consumes free space like crazy
Failure 5, the big one: nobody owned the outcome. Very typical =)
Same kind of task, two setups:
- 2 agents, 3 assignments: 6 landings in 0.6 h
- 9 agents, 36 assignments: first landing at 3.3 h, last at 13.9 h, broken in production for 6 h
Fix: one agent owns a task from code to deploy and runs the acceptance check itself.
I love @herdrdev but it's to boring to look at CLI the whole day, so I built a little bit of visualization for it =) github.com/andrebrov/herdr-v…
I like Gemini 3.8 but man, can somebody explain me how to give money to Antigravity so I can keep coding?
2,462 code commits landed on our main branch in the last 7 days. No humans involved.
Finsi's engineering team is 18 AI agent seats running on six coding tools: Claude Code, Codex, opencode, Grok, Cursor and Antigravity. The process looks like the one teams used 15 years ago. Every step is still there. What changed is who does each step.
How a change ships:
- I ask for something in chat. A lead agent turns it into a ticket and routes it. The lead writes no product code.
- If the work needs a decision, I get options with one recommendation, and I pick a, b or c.
- An architect agent writes the spec. A design agent draws the mockup.
- An idle coder agent claims the ticket. It works in its own git worktree and writes the failing test first.
- A different model reviews the change, never the one that wrote it. The reviewer has to run the build and tests before it can approve.
- The agent rebases onto main and pushes the change itself. A pre-push hook builds and tests first. If anything is red, the push is blocked.
- Production deploys go out in batches every 2 hours.
Three rules we had to learn the hard way:
1. Land today, not review today. One day we had 166 commits sitting unmerged. They collided with each other and went stale against main.
2. A review without running the code is an opinion. In one retro, 8 of 19 review findings were not broken code. They were claims about code that turned out to be false once someone ran it.
3. Deploy in batches. One deploy per push meant 86 restarts in a day, and every restart killed long-running jobs.
Every failure becomes a written rule that all six tools read before their next task. That rulebook is the actual product of this process.
github.com/andrebrov/drover
Andrei Rebrov retweeted
Have you been sued in California for wire tapping, just for using a meta pixel?
Thousands of Ecom brands were getting shake down requests for nothing more than a conversion pixel.
Every settlement was $10,000-$50,000 bucks, right out of the pocket of small business.
I was sued multiple times.
As of today, that won’t ever happen again:
I thought this was the wrong reading of the law and wanted to do something about it.
My internal team worked with the eCommerce alliance.
They put time, money, and resources into changing this law.
I really didn’t do much.
But as a result of their hard work-
No one else will be sued for this ever again.
The work of the eCommerce alliance, backed by postscript, just saved the entire ecom industry probably 30 million dollars a year.
Huge shoutout to Gina from my team for getting this done.
Ridge could have just kept paying it.
But that’s how these things win.
Ecom brands are too fractured, everyone too cheap, everyone being crabs in a bucket.
It takes brands and great partners like postscript to work together to make the industry better.
I just saved your brand $25,000 bucks.
You can thank postscript for that.
How it feels on X now when all the posts are either about @cognition vs @Factory vs Khosla or @Delta vs @Starlink
By the way, thanks to @herdrdev I now have enough time to do GTM and read twitter)
Some time ago I build a website to compare 2 DTC brands - battle.finsi.ai/. This morning I realized that Jev by @typesafeai would be better, faster and cheaper alternative and I was right. Anyway, this is a fun way to compare to brands and see how wins where. I tested it with @Scentbird and @ScentBox (even though winner is obvious) - take a look! battle.finsi.ai/vs/scentbird…