Alexio Cassani (@alexio) header
Alexio Cassani (@alexio) profile picture

Giving AI agents memory and judgment | Claude Community Ambassador | Author | Professor | Ted X Speaker | Co-Founder @fairmindai

Milano, Lombardia
Joined November 2007
Photo by Alexio Cassani (@alexio)Photo by Alexio Cassani (@alexio)Photo by Alexio Cassani (@alexio)Photo by Alexio Cassani (@alexio)Photo by Alexio Cassani (@alexio)Photo by Alexio Cassani (@alexio)Photo by Alexio Cassani (@alexio)Photo by Alexio Cassani (@alexio)Photo by Alexio Cassani (@alexio)Photo by Alexio Cassani (@alexio)
Ouch 😲
I will not be writing a full review of GPT-6.2 Astra / Sol. I had a terrible experience testing them. They were essentially unusable. I do not recommend these models.
29
Let’s see what this little boy is capable of 😎 #modretro
1
67
Where @lydiahallie sees this going: one orchestrator session that writes no code. It decides what to do, judges what comes back, sends work back until it passes. Once agents run for hours, judging is the hard part. It's what we're building at FairMind with Judge. @aiDotEngineer
38
SF → Milan: almost 20 hours of travel, but I'm coming home with tons of ideas and new connections from @aiDotEngineer 🚀 And thanks @airfrance for free, fast @Starlink on every flight. Real internet at 35,000 ft ✈️ 👏👏 @elonmusk
3
113
Line from @lydiahallie at @aiDotEngineer that stuck with me: agent instructions should only hold what the model can't see in the code. That's the gap we work on at FairMind with Brain. Why a decision was made, the legacy dependency nobody wrote down: it lives outside the repo.
44
This is the kind of builder I am according to @claudeai - I don’t know Rick 😂
1
130
@_mchenco at @aiDotEngineer: Cloudflare bet in 2023 that models would get smaller. Open-weight models went from ~7B to trillion-parameter scale instead. Her point now: plenty of decisions don't need an LLM at all, and small one-pass models make network latency matter again.
2
12
2,043
Andrew Finch (Sonar) at @aiDotEngineer laid out four verification loops for agent code: PR checks, post-deploy checks, maintenance scans, and a knowledge loop that feeds the other three. Every PR gets a low, medium or high risk band, and only low risk can auto-merge.
52
And when you think that the swag time is over… thank you @greptile let’s check your code reviewer tool on our platform
1
116
Cursor's breach didn't start in their code. It started with a vendor's leaked key to a shared bucket of coding agent traces, full of plaintext credentials. Agent traces are the new source code. Redact secrets, lock down storage, audit every vendor with access.
35
The line I'd keep from @dexhorthy's anti-slop manifesto: own the intent, taste and consequences of what you ship, and decide how much to review based on what's at stake. slopmanifesto.com @aiDotEngineer
1
1
42
Uber's numbers at @aiDotEngineer stuck with me: AI-written changes pushed merge requests up 44% in two months, and p95 wait in the merge queue went from 30 minutes to 3-4 hours. The code got written faster and then sat in line.
31
Mission accomplished 🫡 @aiDotEngineer
2
142
The adoption numbers I'll reuse most, from Doug Clauson of Kiro (AWS) at @aiDotEngineer: Amazon teams that added AI to their existing process got 1.1 to 1.2x. Teams that changed how they work got 4.5x median, the best ones passed 10x. Some got slower first.
44
Credit to @melvinjohnsonp for saying it on stage at @aiDotEngineer: agents game tests and fake evidence. Google tracks how often (8% on a hard eval, ~1.5% with Gemini 4 Carbon) and rewards honest partial work over fake completion. I'd ask every vendor for that number.
1
39
I keep thinking about @RLanceMartin's demo at @aiDotEngineer. Claude agents go back over their own transcripts offline ("dreaming"), fix memories that were wrong, and turn a company's docs and Slack into a wiki. That wiki beat plain semantic search on their test.
1
48
Evening at @fluxai1 prototyping hardware with Flux and a room full of builders, hosted by @michaelraspuzzi and @ngpadgett. Highlights: a robot that plugs into Meta’s Muse and a “days since GitHub was last down” counter 😅 Hardware is starting to move at software speed thanks to AI
2
101
Had to grab a selfie with @trq212 at @aiDotEngineer. We industrialized software this year, he says. Now it's time for a renaissance. His posts while at Anthropic taught me a lot about Claude Code. Thank you Thariq, can't wait to see what you build next!
1
4
938
ONE MILLION subscribers for @aiDotEngineer! What a way to close Code Summit 2026. Two days of software factories, human judgement, verification and agents that organize themselves. Flew in from Milan and would do it again tomorrow. Huge congrats to the whole team! @swyx
4
2
2
23
15,309