Salvo que estés en el plan max, planificá con Opus y ejecutá con Sonnet 4.6
Si estás en el plan max el effort high debería de ir bien con Opus 4.8 y durarte bastante, dependerá de las tareas y del tamaño del proyecto (si abre 20 agentes en paralelo para algo que le mandás no te va a durar nada)
I found out several people does not understand the key concepts of current LLMs (what the model is capable of, what not, training, embeddings, attention, transformers, RAG, fine-tuning, evals and so on), people heard about those concepts but they don't understand what they are (not talking about ML engineers but other devs or even non-technical people)
So explaining from a no-code experience and with nice visualizations and interactive lessons helped to learn those concepts and understand better the current capabilities of AI
Trabajo en una startup de 5 ingenieros y es tal cual, todos hacen todo y conozco muchas startups nuevas con el mismo mindset: equipos chicos, Product Engineer seniors + AI = se puede escalar muchísimo
Desconozco si realmente escala cuando la empresa crece mucho o si en cierto punto se va a volver a los especialistas para ciertas tareas, pero para crear startups creo que es lo que más está funcionando
Amazing! Considering the new eve observability and the recently released `eve/evals` package, is the idea to expand Vercel to support datasets and iterate/improve with production data the agents?
I wonder how far it could get this agent framework, it seems like just the beginning
@pcshipp One thing is for coding, other for their agents, many startups are using Gemini models for their LLM calls in their apps
Why? Cheaper, good enough models and usually Google has very good startup "free credits" program (better than OpenAI and Anthropic)
@TimJayas The limits are clearly not the same, that's why Cursor tries to push their Composer model
Claude/Codex subscription = highly subsidized
Anyhow, depends on what you do Cursor plan could makes sense
@tomasmalamud@vercel@lapyme_ar Genial! Es más fácil que hacerlo "a mano" con el AI SDK?
Le estás metiendo alguna tool de observabilidad? Me da curiosidad como integraste los botones de feedback (si armás un dataset con los evals para eso o por ahora lo hacés todo más a mano)
@SaraDiscovers Not really, in companies also you could be working on another platform or squashing the commits before merging therefore it will look less than it is
Nowadays even worse with agents commiting directly to main every 2 minutes in side projects, it's just a vanity metric
I've been doing something similar in a more manual way (giving the tools to verify the results by itself then supervise, repeat) and it's great to automate it with built-in loops
The hard part here is the verification, when having a clear number goal is easy, but on building some complex features either I need a good model routing (sometimes I se some model for implementation but I prefer using GPT 5.5 xhigh for adversarial feedback) or I have to step-in to correct what I missed in the original prompt
Aplica bastante bien la frase "el pasto es más verde del otro lado"
La realidad es que todo depende a lo que uno está acostumbrado, la gente que creció en una casa así ve un piso de 50m2 y se siente encerrado, en el otro sentido la gente que creció en un piso ve que tiene todo lejos y la pasa mal
También las edades es un factor clave, de joven se quiere la gran ciudad pero de más grande suele cambiar
@Bhavani_00007 Don't replace it, complement it with Claude, Codex is amazing at reviewing (code and plans), use it with its latest model (GPT 5.5) and xhigh reasoning
In complex projects it'll save you a bunch of time in the review phase
@pcshipp Opening a new chat/session to brainstorm new ideas, doing a deep web research on how to improve what I built, analyzing new Claude skills to experiment
I'd spend more time on giving the AI more tools to auto-verify the work it generates if that's the case
Better linters, better type-checks, connect it to Chrome Devtools MCP so the AI can auto verify the work end to end, connect it to any provider you have (observability for example), give it visibility to your GitHub (and CI) using gh CLI to auto fix CI issues, give it a way to connect to your local DB and let it debug itself with the psql CLI for example
@nextjs Amazing news, thanks for listen the feedback, I'll try it for sure
And nice move to smoothly add the new skills for it (nice complement to the normal codemods on a new version!)
AI SDK 7 is now available.
Introducing: reasoning control, agent-level tool approval, tool and runtime context, file and skill uploads, MCP Apps, durable workflows, terminal UI, sandbox support, harness integrations, telemetry, lifecycle events, and more.