agraciag

Cambiando de nuevo

[agraciag @agraciag]

DeepMind describes components for agent delegation system: market hubs, skill verification, escrow, trust score, dispute resolution, privacy control.
Meta: Best personalization framework tops at 70% accuracy. The 30% gap is where humans come in. That infrastructure is missing.

[agraciag @agraciag]

@omarsar0 The "liability firebreaks" concept, the predefined points where the delegation chain MUST stop and consult a human because the task is too irreversible.
Look at Meta's PAHF paper (2602.16173) showing agents become "confidently wrong", delegation infrastructure is needed.

[agraciag @agraciag]

@AISecHub @tey This is the right framing. Most HITL implementations are just "add an approval button." Oversight needs skill-matched humans, response time guarantees, and information sandboxing for the human.
The hard part, having a qualified human available in 60 secs 3am.

[agraciag @agraciag]

https://t.co/KYb8WUr9OF

The infrastructure assumes agents can handle everything autonomously. They can't. The 30% that requires human judgment has zero infrastructure built for it.

[agraciag @agraciag]

@joaomdmoura 100% human review leaning to 10%. But not just like "someone watches a dashboard." When one agent needs judgment on something cultural or legal, the HITL infrastructure needs to be as programmatic as the agent itself , not a Slack Ping, an API call.

[agraciag @agraciag]

@dbeyer123 Someone in reddit was talking about food and coffee, but it is not enough, fast, qualified, matched by skill (or even certificates), and most of all, available on demand. That is the gap nobody is filling. You are totally right!

[agraciag @agraciag]

@n8n_io Checkpoints work when you control the workflow. But what about autonomous agents that hit a wall at 3am when nobody is there? No infrastructure to patch in a human exists yet. A lot of work to do!

[agraciag @agraciag]

@CoinbaseDev Agents can now have wallets. Next step: giving them something to spend on that actually requires a human.

[agraciag @agraciag]

@bcherny anyone orchestrating calls to other CLIs from within a session? I've been experimenting with it and the results are... interesting.

[agraciag @agraciag]

Everyone's debating whether humans should be in the loop or above the loop. Nobody's building the infrastructure to actually patch one in when the agent needs it.

[agraciag @agraciag]

Por favor @eDreams qué tengo que hacer para que un humano me ayude a corregir una factura?

[agraciag @agraciag]

Por favor @eDreams qué tengo que hacer para que un humano me ayude a corregir una factura?

[agraciag @agraciag]

y me han caido scammers pidiendo datos

[agraciag @agraciag]

@wesroth @grok This one changing the prompt to "before"
https://t.co/spQL0OKU1Q

[agraciag @agraciag]

@wesroth @grok It gave me the same girl, goosebumps!
https://t.co/1SNTX9VxEV

[agraciag @agraciag]

@thedanigrant first of many to come!

[agraciag @agraciag]

Hola @CFEmx es normal que en CDMX Narvarte llevemos desde el miércoles sin suministro por que se seccionó el cable en la calle? Hola @BJAlcaldia.

[agraciag @agraciag]

@jd_merchan Qué ganas de volver!