Dangel | Galaxyhub Labs

Building @GalaxyhubAI | @Voightxyz
| x3 Hackathon Winner
| Member of @LaFamilia_so
| JET Pilot member @JET_EUROPA_

[Dangel | Galaxyhub Labs @seenfinity]

🇪🇸 SOLANA, NOS VEMOS EN MADRID.

@LaFamilia_so presenta: El Encuentro.

Un día para reunir a los builders, founders y comunidades que están construyendo el ecosistema Solana en España.

Charlas, paneles y networking. Pero, sobre todo, la oportunidad de ponerle cara a la gente con la que compartimos este ecosistema.

📍 Madrid
📅 Lunes, 26 de octubre · 10:00–22:00
🎟️ Entrada gratuita - Plazas limitadas

🧵👇 ¿Nos vemos allí?

[Dangel | Galaxyhub Labs @seenfinity]

@eclips3_co @Lid_onchain Bullish 🙌

[Dangel | Galaxyhub Labs @seenfinity]

@albertofdez20 @eclips3_co 🔥🔥🔥

[Alberto Fernández @albertofdez20]

communities that earn a cut of what they grow stick around. glad to be building this with @eclips3_co

[Triton One 🌊🌋 @triton_one]

If you're in Madrid on Monday, stop into El Encuentro 📷with @LaFamilia_so, catch a talk on Alpenglow by one of our engineers, and win some free Triton RPC!

https://t.co/5YF0KkhI4J

[Dangel | Galaxyhub Labs @seenfinity]

@triton_one @LaFamilia_so Niceee 🔥

[Dangel | Galaxyhub Labs @seenfinity]

@LaFamilia_so Activo con MetaÑOL!

[La Familia @LaFamilia_so]

MetaÑOL EP 13 https://t.co/n4XOYSC6ZM

[Dangel | Galaxyhub Labs @seenfinity]

@sh_reya From building agent observability: the messiest traces are always at the human↔agent handoff, not inside the agent loop. Systems designed for that boundary would be a real unlock.

[Dangel | Galaxyhub Labs @seenfinity]

@useDiversiFi @solana @LaunchOnSF Bullish ! 🔥

[Dangel | Galaxyhub Labs @seenfinity]

@nayibbukele

[Dangel | Galaxyhub Labs @seenfinity]

@amasad Memory is where agents fail quietly: a stale fact retrieved with confidence looks exactly like a correct one in the output. How are you thinking about provenance and expiry, so a user can see which memory drove a given action and correct it?

[Dangel | Galaxyhub Labs @seenfinity]

@jasonlk If the agent is the buyer, the hard question shifts from "did it pick the right vendor" to "can you reconstruct why it did". When yours churned vendors in 12 hours, did you get a decision log with the criteria it weighed, or did you have to reverse-engineer it from the invoices?

[Alberto Fernández @albertofdez20]

the @Lid_onchain builder is ready. you describe what you want to sell and it builds the page for you. spent tonight testing it with @Darwinerc92 and @_cxalem

[Dangel | Galaxyhub Labs @seenfinity]

@albertofdez20 @Lid_onchain @Darwinerc92 @_cxalem Nice team

[Dangel | Galaxyhub Labs @seenfinity]

Turning successful agent runs into evals is the part I keep running into while building agent observability, because "success" is almost never labeled at the moment the trace is captured. How do you decide a run actually succeeded: merged PR, green tests, or a later check that nothing got reverted within N days?

[Dangel | Galaxyhub Labs @seenfinity]

@a1lon9 Nice

[Dangel | Galaxyhub Labs @seenfinity]

That refund case is the perfect example: every step "succeeded", the outcome was wrong. Feels like the fix is encoding intent as invariants (refund.customer_id == order.customer_id) and checking them at the write, not after. How did you catch it, alerts on the invariant or a customer complaint?

[JET Europa @JET_EUROPA_]

Another chance to win $100. 🏁

Yesterday, we crowned our first JET Nitro winner.

For this second round, we only care about one thing:

Who can generate the most trading volume on JTX?

Enter below. ⬇️

[Dangel | Galaxyhub Labs @seenfinity]

@aryakanna9 @hwchase17 Agreed, traces only tell you what the agent claimed it did. Pairing each run with a state diff (DB rows, files, external API side effects) is what catches the "looks fine in the log, wrong in prod" cases. Do you snapshot before/after per run, or diff against an expected state?

[Dangel | Galaxyhub Labs @seenfinity]

@GoldmanSachs @GS10KSmallBiz Training is the easy part. The hard part for a 10-person business is that nobody on staff has time to fix an automation when it breaks, so "tools and support" has to include someone who owns it after go-live. Will the Institute fund implementation, or only training?