Building Agentic Data Intelligence (ADI). Replacing BI with autonomous data agents. Data engineering • AI • Enterprise
Warsaw, Poland
Joined October 2007
- Tweets589
- Following266
- Followers1K
- Likes512
The most discussed topic during crypto conferences a few years ago was tokenomics. The most discussed topic during AI conferences today is tokenomics.
The movie has a sequel: oracle poisoning ☠️🤖
Last time the system fooled itself. This time someone does it on purpose.
Researchers are already documenting it in agent stacks: “the better the agent reasons, the more convincingly it propagates the attacker’s false premises” 😩
That’s the job. Not compliance paperwork but an anchor with manipulation resistance.
Building it at @agentic_data ⚓️
Paper + full argument 👇
Oracle poisoning paper:
arxiv.org/abs/2605.09822
The longer argument on why agent graphs need an anchor: governedagent.substack.com
mierzwik retweeted
Nadciąga fala awantur w matematyce. Modele AI są już wystarczająco silne, by wskazywać błędy w opublikowanych dowodach, więc będzie wesoło…
The answer was never "add another smart contract that checks the first one" 🚫
It was governed external anchors: whitelisted oracle feeds with owners, a security module, delays, etc ✅
The fix wasn't more loops. It was an anchor the loops couldn't tune ⚓️
Enterprise data needs the same thing before agents run on it: semantic definitions with owners, enforced OUTSIDE the loops that could easily game them 🫡
That's what we’re building at @agentic_data ⚓️
agentic-data.com
Ask an agent for FX exposure, and it writes SUM(fx_rate). The query runs. The judge model approves it. The evals score it reliable. But the number is garbage. A rate isn't additive, and nothing in that stack knows what the column means.
More loops won't fix this.
An anchor will.
So maybe we retire "the AI went rogue." It did exactly what its environment allowed, and the environment allowed catastrophe.
Almost good news, really: you can't patch an agent's judgment, but you can fix where the decision gets made.
I wrote the full second-by-second post-mortem, plus the blast-radius audit and an incident-response card you can lift:
open.substack.com/pub/govern…
This failure class has a name, too. That's the next one.
Sources:
• PocketOS: theregister.com/software/202…
• Replit: fortune.com/2025/07/23/ai-co…
(The agent was Cursor running Claude Opus 4.6.)
In April, an AI coding agent wiped a startup's entire production database in 9 seconds, and the backups with it, since they sat in the same volume.
The founder, @lifeofjer, posted the whole story.
But the agent isn't really the story. 🧵
