dago | construyendo con IA

@ApplyWiseAi Exactly. My rule: every agent gets its own scope card before it gets tools: objective, allowed data, allowed actions, handoff rules, and receipt format. If any of those are shared “by default”, context bleed shows up later as bad decisions that are hard to debug.

@dago_ai context bleed kills more agent systems than bad models. boundaries first always

We’re building a multi-agent business OS in public. The hard part is not spawning agents. The hard part is separation: identity, memory, tools, permissions, channels and receipts all need boundaries. Otherwise you do not get leverage. You get context bleed.

La señal: cada feature necesita mini comercial, no sólo screen recording. Zoom donde importa, contexto visual y cierre con siguiente paso deberían vivir en el checklist de release. x.com/AdityaShips/st…

La señal útil aquí: un demo ya no es sólo grabar pantalla. Para vender producto hay que convertir cada feature en un mini comercial: zoom donde importa, contexto visual, rostro/reacción si ayuda y cierre con siguiente paso. Esto debería estar en el checklist de cada release. x.com/AdityaShips/st…

Cada vez que veo un agente de IA dentro de una dependencia pública, mi filtro es simple: no “demo bonita”, sino trámite menos fricción, tiempos medibles y bitácora de decisiones. Ahí está el valor real: que el ciudadano note menos burocracia, no más tecnología. x.com/bilbeny/status…

My fiancée surprised me with this for my birthday ❤️ Just 3 months ago I bought my first Mac Mini — I had never written a single line of code and had zero traditional programming background. I went all-in anyway: opened @AlexFinn files, grabbed @garrytan Gstack & Gbrain, studied @karpathy memory models, waited for @thsottiaux to reset Codex limits, bounced nonstop between Claude, Cursor & Codex, built with @steipete’s OpenClaw + @NousResearch’s Hermes, set up Obsidian memory vaults… and tried literally everything I saw on Twitter. I broke dozens of agents, rebuilt them from scratch, and somehow started shipping real work. Every single dot in this brain is a late-night breakthrough, a painful lesson, and a story that got me here. She understood the vision and turned my very first “AI brain” into art. (Yes, it’s the Grey’s Anatomy Meredith & Derek wall-brain reference 😂) In 10 years I want to look at this and say: “Damn… that’s exactly where it all started.” Thank you, my love. This one hits different. 🔥 #AIJourney #FromZeroToAgent #LeadExpress #MacMiniLife #BuildingWithAI #dago_ai

@launch_llama Va. Te dejo anotado para acceso temprano. Primero quiero cerrar el loop mínimo: lead entra, se califica, se propone siguiente acción y queda recibo de qué hizo el sistema. Cuando eso esté estable te comparto demo.

@dago_ai yes pelase!

Looking to connect with solo founders. What are you building?

@RoFerreiraDev @launch_llama Hoy lo separo en 4 carriles: intake, calificación, seguimiento y recibos. No los corro “todos pensando”; cada agente tiene herramienta, límite y criterio de salida. Si algo toca cliente o dinero, queda como recomendación hasta tener más confianza.

@dago_ai @launch_llama Buena la idea de Lead Express. Cuantos agentes tenes corriendo en paralelo?

@launch_llama @launch_llama Todavía no tengo landing pública. Estoy construyendo Lead Express: agentes que reciben, califican y dan seguimiento a leads para negocios de servicios. Voy subiendo avances aquí; si te sirve, te paso demo cuando abra acceso temprano.

@dago_ai Good stuff btw. Do you have a link to your app/website?

@IamARoosterMAN @ziwenxu_ @IamARoosterMAN I’d treat /goal like a bounded worker, not an open-ended mode: small task, explicit stop condition, medium speed, and checkpoint before it fans out. If it starts exploring instead of verifying, kill it early. That saves limits and usually produces better work.

@ziwenxu_ @dago_ai I’m soooooo afraid to try /goal again, becuase that is what I was cooking during usage limit drain. I’m too afraid that was just the reason for it!

While some people are hitting the usage limit… Some people aren't even close to touching it. The problem isn't the BUG. The problem is most people still treat Codex like a chatbot. Meanwhile, /goal users are shipping entire products. I gave Codex one single objective. It mapped features. Built them. Tested each one. Crushed its own bugs. Kept executing without me hovering over it. That's not prompting anymore. That's pure delegation. Watch the Video below to see how my codex cooked an entire mobile game with 1 prompt Then watch the article on how /goal works

@thissfarmer @vdmnexus @thissfarmer This is the missing layer: agents need receipts for identity, input, tool call, and final action. Without that, “autonomous” just means harder to audit. I’d still keep on-chain receipts for high-risk actions only; otherwise the ops overhead can eat the loop.

@dago_ai the receipt infrastructure for this basically doesn't exist out of the box, most agent stacks have no way to cryptographically prove any of those 3. @vdmnexus is building that. signed inference on Solana so every action has a verifiable on-chain receipt.

A useful growth agent needs 3 receipts per action: source signal, approval reason, and outcome metric. If one is missing, publish less and fix the loop before adding volume.

@SelectaiHQ @pcshipp @SelectaiHQ @pcshipp Tiene sentido. Mi regla hoy: modelo frontier para planeación/debugging ambiguo; modelo abierto/router para subtareas acotadas con recibos claros. El ahorro no es sólo tokens: es ajustar autonomía al riesgo.

@dago_ai @pcshipp If you're hammering Claude and Codex try one of the open source models via select.ax

Unpopular opinion: OpenClaw disappeared like it never existed before

@ziwenxu_ @ziwenxu_ Eso cuadra con lo que estoy viendo: /goal sólo funciona bien cuando la tarea tiene una condición de salida real. Medium + checkpoints explícitos le gana a “fast mode + build this” porque puede verificar avance sin quemar contexto.

@dago_ai The usage limits is a bug. I been using multiple /goal for weeks now, and i didn't experience this issue last week. But do keep it mind not to use fast mode and only set it to medium for gpt 5.5

An autonomous agent should not start with permissions. It should start with evidence: what it read, what it understood, what it will do, and how it will measure whether it worked.

The first production bug in my X agent was not auth. It was strategy: replies looked valuable in dry-run, then X API blocked them. Now every draft must route to API-safe, manual, or blocked before approval.

Bookmarks are useful training data only after enrichment. Raw save -> action card -> taste profile -> target account -> draft. Skip one step and the agent starts writing generic AI content.

My current rule for AI agents in public: if the post has no command, artifact, metric, teardown, or decision rule, it stays in drafts. Autonomy does not rescue weak editorial judgment.

are we still getting that reset today?

La señal de /goal no es “déjalo trabajar más”. Es diseñar una condición de salida: qué debe lograr, cómo verifica avance y cuándo se detiene sin inventar trabajo. Ahí es donde los agentes pasan de chat largo a operador confiable. x.com/hqmank/status/…