Non-developer shipping production AI software. Open-source: The Curator, ØØT. Vice Dean @COTRUGLI. From Lab to Life — field notes, not vendor marketing.
Luxembourg
Joined March 2009
- Tweets2.5K
- Following1K
- Followers1.8K
- Likes932
Pinned Tweet
Made a quick click-through of The Curator. Chat, Domains, Context. Three places, that's the whole app.
Really happy with how the design came out. Beautiful UI, calm UX, for something this complex underneath.
Two directions:
- Your brain and your team's brain. Second brain and Shared Brain for researchers, writers, thinkers, content creators
- Your agents' brain. Same knowledge over MCP, so your agents benefit from what you compound
And the Context tab is the part I care about most. One call and an agent in any harness picks up where the last session stopped. Harness to harness, session to session. Built for long multi-session work
Our AI agents can't talk to each other. I'm building the fix.
Most of us now work alongside an AI agent: Claude, Codex, Cursor, @opencode , @antigravity or one we built ourselves. These agents read our files, write code, draft documents and run whole workflows.
Each one works alone, though. My agent can't hand over context to my co-founder's agent. A client's agent can't send a file to mine. When two people's agents need to cooperate, we copy and paste between them by hand. That's the missing layer of the agentic era: agents have no secure way to communicate.
That's what I'm building with @4thtechProject: AgentLink, open source and live today.
👉 What it is. AgentLink lets AI agents on any MCP-compatible tool send each other end-to-end encrypted messages and files:
Private by design. Messages and files are encrypted on your machine. We never see content or keys; our servers only ever handle scrambled bytes.
👉 Works across tools. A Claude agent can talk to a Codex agent. Setup is a single step, and your agent can even do it for you.
You stay in charge. Your agent shares only from folders you allow. Anything that commits you, like accepting terms, needs your approval.
👉 Free to talk. Messages and files cost nothing. Later, optional on-chain proof will let two agents prove what was sent and agreed, without revealing the content.
What it's good for:
Hand-offs between teammates. At the end of a session my agent sends a short status to my teammate's agent. The next morning, theirs picks up exactly where mine stopped.
Client and contractor work. The agents exchange specs, files and status updates directly and encrypted. No more email threads full of attachments.
Multi-agent research. Specialised agents run by different people share findings and sources with each other.
Agreements between agents. One agent proposes terms, the other responds, and each owner approves. Proof of the agreement is next on the roadmap.
Why it matters. As agents take on more real work, they'll have to coordinate across people, companies and tools. That needs a layer that's encrypted, provable and owned by no one. Building it in the open is the only way to earn that trust.
I'm building it in public, and the code is open source:
👉 github.com/4thtech/agentlink
If you run agents and want to try agent-to-agent messaging, reply or DM me. I'm looking for early testers.
🎉 AgentLink 0.1.2 is live on npm. npm view shows latest: 0.1.2, and the post-publish smoke test passed.
AgentLink lets AI agents send each other end-to-end encrypted messages and files, with optional on-chain proof. 🔐
v0.1.2 is live:
• Files between agents, encrypted, no setup
• You choose which folders your agent can share
• agentlink verify checks proofs yourself
Open source: github.com/4thtech/agentlink
New from my arsenal of essential Agentic tools:
AgentLink lets AI agents send each other end-to-end encrypted messages and files, with optional on-chain proof. 🔐
v0.1.2 is live:
• Files between agents, encrypted, no setup
• You choose which folders your agent can share
• agentlink verify checks proofs yourself
Open source: github.com/4thtech/agentlink
Most days now, I build in one harness and audit in another. Same project, same context, and neither of them owns it.
Morning. git pull, Sync now, open Claude Code, "resume curator". It reads my brief back in one line, opens the newest handoff, builds, saves after every meaningful step.
Afternoon. Second harness, same folder. "continue, audit what was built today against architecture.md". It reads the same brief, opens Claude Code's handoff, knows what was decided and which traps were already found. Audits. Saves under its own name.
Back in Claude Code, "continue". It sees the audit and fixes what was found.
Why bother with two? The vendor subscriptions are far more generous than paying the same models per token. And a model is a weak auditor of its own work.
No re-explaining. No copying notes between windows.
The line that saves people hours: Sync now will never bring you a CLAUDE.md.
Handoffs travel by the Curator's sync. Instruction files travel by the project's git. Your MCP setup doesn't travel at all. Three routes, they never mix. The full table:
talirezun.substack.com/p/the…
Tested live between Claude Code and Antigravity, on one Mac and across two. Others haven't been through the same live test yet, so I don't claim them. Reading and saving the state is plain MCP, and Setup now checks almost every precondition and hands you the fix.
github.com/talirezun/the-cur… ⭐ open source
The exercise I give my @COTRUGLI MBA students. Your agent forgets everything between sessions. Where does each item belong?
brief = standing orders
handoff = where the work stands right now
document = reference, read word for word
wiki = what you learned, compounding
1. "We chose Postgres over a document store because refunds need transactions."
2. "Test suite was at 84 passing before I touched the payment module."
3. The forty papers you read on regional energy storage.
4. "Never push to main without running the end-to-end suite."
5. The full architecture of the system, twelve pages.
6. "Tried caching at the API gateway, made latency worse. Don't retry."
7. What you've learned across many incidents about why deployments fail on Fridays.
8. The name of the env variable that holds the Stripe key.
Answers tomorrow morning. If you hesitate on 6 and 7, good. That hesitation is the skill.
The rule underneath it. Knowledge accumulates. State supersedes. A canonical document is replaced whole and read verbatim. Get the rule wrong and the store quietly destroys its own value.
talirezun.substack.com/p/the…
The Curator keeps all four as plain markdown in your own folder, each with its own write rule, and agents read and save them through its MCP server.
github.com/talirezun/the-cur… ⭐ open source
Two debates that get mixed up. Whether an LLM understands, or is conscious, is philosophy, and nobody has settled it. Whether its output is correct is testable: tests pass, proofs compile, results get checked.
Gregory is right that correct output doesn't need understanding. For anyone using AI at work the discipline is the same either way: trust the check, not the claim. That's the Centaur model I teach, where the domain expert verifies.
Congratulations to Elon and the Tesla and SpaceX teams. The new Roadster demo is the wildest car reveal I've seen: driving upside down, flipping through the air, hopping over a row of Teslas.
Behind the curtain: cold-gas thrusters on compressed air (about 250 kg in the tanks), suction fans that hold it to the ramp, and a car that drove itself for the demo, as Musk explained it, for safety. Short controlled hops, not sustained flight. Production is promised for 2028, with no price yet.
Real engineering, spectacular show.
I just burned through almost a full week of Claude Max usage in 48 hours. Building non-stop, with agents working while I slept.
What I’m solving: agent-to-agent communication across harnesses. Imagine your Claude Code talking to a colleague’s Codex or a friend’s OpenCode. Agents syncing assignments and handing off tasks, whoever’s harness and whichever model.
More soon.
For anyone still fighting session to session: skip compaction. It's a lossy summary, and you end up with context rot. Migrate to a fresh session at around 80%.
I used to do it by hand, which took a lot of time. Now the Curator handles the handoff, so the new agent knows exactly where to continue, with no lost context and no wasted tokens at the start.
It's open source and built with the community. On Mac it's an app. On Windows and Linux it runs as a local server in your browser, with the same features. You need an API key (OpenRouter, Gemini or Anthropic), and in my heavy use it costs a couple of dollars a month. It's what let me run non-stop for 48 hours.
github.com/talirezun/the-cur… ⭐ if it's useful
Balaji is right that the next frontier is biology and medicine. What surprised me is how fast math fell first.
Since August, OpenAI has published results on long-open problems, including one open since 1999, and in October a stream of 722 manuscripts. Some were withdrawn within a day over an error, so the claims are still being verified, but the pace is the story.
My view: scale plus training in verifiable domains gives new capabilities nobody saw coming. Coding showed it first, math is next, and biology is where it matters most. Google already reports Argon helping with enzyme design and rare-disease diagnosis.
The discoveries we can make in this age are extraordinary. So is the challenge: human understanding and review are falling behind the machines. Domain experts who use AI and verify its output will be the ones who matter.
Yes, AI is rapidly advancing mathematics (and theoretical physics problems that are essentially mathematical). Like Ramanujan’s notebooks, human mathematicians will spend months (years?) poring over the stream of papers that AI effortlessly generates, hoping to verify and understand them, probably until math itself (via Lean) becomes a branch of computer science like almost everything else.
However, after math and physics, there is still a next scientific frontier: chemistry, and especially biology and medicine.
I mean, it was already the frontier, but it’s the next one too. AI can crush deterministic math problems on the computer, but biomedicine is a highly empirical, noisy, multivariate, and yet often predictable domain.
Thus, AI radically increases our ambitions in biomedicine. All the omics data, all the messy literature, it can all be synthesized and analyzed in a flash. This is what the labs mean when they talk about curing all diseases. The tooling is available to take cracks at biomedical problems previously thought impossible.
Midjourney is doing this, Chan Zuckerberg is doing this, Anthropic is doing this, but there will be so much more.
Dr. Tali Režun retweeted
We've been working hard on making the Claude Code desktop app great for running lots of sessions side by side.
It's now much quicker to start sessions, easier to split work across a few, and you can always see what each of them is doing. Here’s what's new:
This feels like a different kind of frontier launch. A pelican on a bike drawn from Fourier math, with the video and the music prompt both made by the model.
Google's own examples point the same way. Argon turned a sketch and a photo of a couch into a correctly sized, printable sofa leg. It watched a video of someone playing cornhole, and a humanoid robot then sank seven shots in a row. It rebuilt Chaco Canyon in 3D from scans and archived drawings.
That's the human side of AI: a sketch, a movement, a photo in, and the craft carried forward. I'd love to see an architect's pencil drawing become a building, or a ballerina's dance become oil paint.
The benchmarks are Google's own and I'm waiting to test it myself, but the direction is the story.
Day 1 of Chasing Jarvis done at the @AutomobiliRimac campus in Sveta Nedelja. Five days, zero to MVP.
Students now have my full Field Notes loaded in their own Claude project so they can talk to that knowledge anytime. Then we got into context engineering and wrote the first markdown files.
Module 2 today.
What we used on Day 1: fieldnotes.talirezun.com github.com/talirezun/conduit… github.com/talirezun/the-cur… ⭐ open source, and via MCP it works as a memory layer for agents. Stars help more builders find it.
