Saturday guide: the week agents stopped being a cloud luxury

AI agents leave the academic-paper shelf and enter the reality of deployments with no fixed cost. Meanwhile, models get more efficient, cost less, and your IDEs gain intelligence that knows the backlog. Sit down and read.

To do: the tutorial of the week

[How to run the OpenAI Agents API on Vercel without keeping a VM running](/front-end/como-rodar-a-openai-agents-api-no-vercel-sem-manter-vm-ligada) — by Carina Ferreira.

The whole path to getting an agent with session memory running on Vercel, without paying for an idle virtual machine.

To read: three from the week

• [SWE-2, from Cognition, promises to rival top-tier models at a quarter of the cost](/noticia/swe-2-da-cognition-promete-rivalizar-com-modelos-de-ponta-a-um-quarto-do-custo) — SWE-2's coding model reaches a cost level that starts to question where the real gain is against the benchmarks. (Redação iMasters)
• [DeepSeek v4.1-Flash separates prefill and decode and drives down the cost per task](/inteligencia-artificial/deepseek-v4-1-flash-separa-prefill-e-decode-e-derruba-o-custo-por-tarefa) — DeepSeek proves it's possible to make inference cheap even with the KV cache being latency's biggest headache. (Alan Andrade)
• [Copilot gets Jira integration and automatic model routing in the CLI](/back-end/copilot-ganha-integracao-com-jira-e-roteamento-automatico-de-modelos-no-cli) — Copilot finally knows how to work with the systems you already use, not just inside the editor. (Bisneto Braga)

To reply: the conversation of the week

On the iMasters Forum: BC provider authentication issues when using JCE in Java Swing applications, started by Fabio Barros, with 3 replies so far.

The weekend is good for this: tell us on the forum what you're working on.