Daniel Alonso

Co-founder at @DarumaLabsAI

[Daniel Alonso @apanoia]

Since 2 Aug 2026 an AI that talks to people in the EU has to say it is an AI. Most agent frameworks have no field where you'd write that down.

I spent the last weeks making the EU AI Act something jaato can express, check and prove. A thread follows. https://t.co/p5AooDq99E

[Daniel Alonso @apanoia]

@YMaxwellHayes Indeed, disclosure is framework controlled, not up to the model to decide.

[Daniel Alonso @apanoia]

Since 2 Aug 2026 an AI that talks to people in the EU has to say it is an AI. Most agent frameworks have no field where you'd write that down.

I spent the last weeks making the EU AI Act something jaato can express, check and prove. A thread follows. https://t.co/p5AooDq99E

[Daniel Alonso @apanoia]

@Verimand The article in question is more about the EU AI Act, mandating "transparency and maintenance under human control" of agentic based applications, not so much about RBAC, perms mgmt in general, which anyways are also defined, enforced and tracked by config knobs in Jaato's profiles

[Daniel Alonso @apanoia]

@freesolodev Closed, and thanks - it was a real gap. jaato-server 0.17.0: an announcement
record in the chained ledger with the text as delivered, the client type, a
declared locale and the provider/model. Suppressed is recorded as suppressed,
not absent. https://t.co/Fbq9FeLU3i

[Daniel Alonso @apanoia]

@freesolodev You're right, and it's a gap. Today the ledger binds model identity and user to each round trip, and the announcement goes out as an event before the first turn, but the announcement text, channel and locale are not written into the chained record. That is the next one to build.

[Daniel Alonso @apanoia]

@YMaxwellHayes Both, and the protocol half is the primary one. Every output event carries a framework-filled source (model, tool, system); the disclosure announcement rides it as system before the first turn, so a client renders it. The prompt piece is only the fallback for when a person asks.

[Daniel Alonso @apanoia]

@anrayama It already does. The warning names the field, regulatory.interacts_with_persons, and the Article, 50(1), and says what true and false each mean. Under risk_class: high the same finding is an error. One edit, as you say.

[Daniel Alonso @apanoia]

@anrayama Close, with one correction: a missing disclosure declaration is a warning by default and an error only under risk_class: high. The framework never infers the class, so it cannot fail a build on a determination nobody made. Absent is undeclared, not minimal.

[Daniel Alonso @apanoia]

@FReza1984 By not writing the policy twice. The explain pages and the Annex IV dossier are computed from the code that enforces each control, not restated beside it. If the enforcer changes, the evidence changes with it, and a guard fails CI when a documented control has no enforcer.

[Daniel Alonso @apanoia]

@mitansh_j07 Thanks. That is the bet: a key the validator reads cannot be forgotten, and a ledger you can verify cannot be merely asserted. Both are in the repo if you want to try to break them.

[Daniel Alonso @apanoia]

Run agents in Europe and need to land this? That is our work at @DarumaLabsAI: new agents on jaato, existing deployments brought up to these controls, or agents on another stack moved onto jaato. Your counsel makes the legal call; we do the engineering.

[email protected]

[Daniel Alonso @apanoia]

As far as I know, jaato is the first agentic framework to support these requirements natively. What it deliberately doesn't do is in the article too, and I'd like to hear what I got wrong.

Manual: https://t.co/ho6pZojgwu

[Daniel Alonso @apanoia]

@samueljmcd https://t.co/rAClP7Vq1J

[Daniel Alonso @apanoia]

@GordoDan_ Esa no es la cuenta oficial de la artista española Rosalia, dejad de inventar

[Daniel Alonso @apanoia]

@GordoDan_ Chicos, que quien compartió ese video es una cuenta llamada "La Rosalia", fíjense bien en la foto que aparece al principio, no es de la artista. La cuenta de la cantante se llama "rosalia.vt LA ROSALIA", son cuentas diferentes.

[Daniel Alonso @apanoia]

@dulski_ukasz @Clint_Davey1 Not "of", but "or" ;-)