Ir para o conteúdo

title: "(short descriptive title max 10 words)" url: "https://x.com/i/status/2059993771409015076" collection: ai tags: [(3 to 5 short thematic tags, no file names)] tools: [(only names of real software, languages, cloud services, APIs. No file names, abstract concepts, generic terms. If none, write "Nenhuma")] category: (one of: DevOps, Frontend, Backend, IA, Mobile, Data, Design, Negocios, Outro) date: 2026-05-29

(short descriptive title max 10 words)

(rewritten main content as markdown article following rules)

We need to parse the provided content: It's a tweet about Microsoft open-sourcing SkillOpt, a framework for training agent skills like neural networks; skillOpt treats a markdown file as trainable parameter of a frozen LLM agent; uses optimization discipline: learning rates, validation gates, batch sizes, epoch schedules; details of workflow; mentions Hermes Agent, GEPA, etc. Provide links to paper and GitHub.

We need to avoid inventing info; use only given content.

We need to produce title <=10 words.

Potential title: "Microsoft abre o código do SkillOpt para treinar agentes LLM". Count words: Microsoft(1) abre2 o3 código4 do5 SkillOpt6 para7 treinar8 agentes9 LLM10. That's 10 words exactly. Good.

Resumo: 2-3 paragraphs describing what the post addresses: talk about Microsoft open-sourcing SkillOpt, description of framework, analogy with neural network training, workflow, mention Hermes Agent, GEPA, links.

Ferramentas: Need to list real software/frameworks etc mentioned: SkillOpt (framework), Hermes Agent (maybe a system), GEPA (optimization loop), maybe arXiv (not a tool), GitHub (service), maybe LLM (not a tool). The instruction: only names of softwares, frameworks, languages, services cloud or APIs real. So we can list: SkillOpt, Hermes Agent, GEPA, GitHub. Possibly also arXiv? It's a repository, not exactly software/service but could be considered a service. Safer to include GitHub, maybe also arXiv as a service? It's a preprint server. Might be okay. But we should avoid abstract concepts. Let's list: SkillOpt, Hermes Agent, GEPA, GitHub. If we want to be safe, we can also include "arXiv". But arXiv is a service. I'll include it.

Category: This is about AI/agents, so category IA.

Tags: 3-5 short thematic tags: skill-opt, agentes-llm, otimização, código-aberto, treinamento. Ensure no file names.

Now corpo: Need to rewrite content as markdown article with:

  • Start with a paragraph of context 2-3 lines explaining what it is.
  • Use H2 for sections, never H1.
  • Use admonitions (note, tip, warning) with proper indentation (4 spaces).
  • Use bullet lists with "-".
  • Use markdown tables when appropriate.
  • Code blocks with language.
  • Important technical terms in negrito first occurrence.
  • End with section ## Pontos-chave containing 3-5 bullets.
  • Do not invent info.
  • Do not include original link.

We need to include the main content: description of SkillOpt, analogy, workflow, mention Hermes Agent and GEPA, links to paper and GitHub.

We need to highlight important terms in bold first occurrence: e.g., SkillOpt, Hermes Agent, GEPA, LLM, markdown, learning rates, etc.

We must not repeat the title. No H1.

We'll structure:

Visão geral

Paragraph context.

Como o SkillOpt funciona

Explain analogy, workflow.

Comparação com Hermes Agent

Maybe a table comparing SkillOpt vs Hermes Agent.

Pontos-chave

Add admonitions where appropriate: maybe a note about benefits, tip about usage, warning about limitations.

Let's craft.

We need to ensure we don't invent info: only use given content. The given content includes: Microsoft open-sourced SkillOpt; treats plain markdown file as trainable parameter of frozen LLM agent; applies same optimization discipline used in weight training: learning rates, validation gates, batch sizes, epoch schedules; analogy maps precisely; skill document is parameter; trajectory-derived edits are gradient direction; edit budget is learning rate; held-out split is validation check; How it works: frozen model runs tasks with current skill and produces scored trajectories; separate optimizer model analyzes failures in minibatches, proposes structured add/delete/replace edits, ranks them under budget cap; if candidate skill improves performance on held-out split, edit accepted; else rejected and stored to avoid repeating; deployed output is single best_skill.md file, typically 300-2000 tokens; no weight changes, no extra inference-time calls; learned rules compact and readable; read like rules a thoughtful engineer would write after a day with benchmark, except discovered automatically; Learn more: Paper: arxiv.org/abs/2605.23904; GitHub: github.com/microsoft/Skil (note truncated? Actually URL: github.com/microsoft/Skil maybe missing "lOpt"? but we keep as given). SkillOpt isn't first; Hermes Agent independently built same idea through skill_manage, Curator, and optimization loop GEPA that scores, mutates, promotes skill documents across runs; Two teams, different architectures, same conclusion: skill file is highest-leverage thing to optimize in frozen-model agent; Author wrote deep dive on Hermes agent.

We need to not invent info beyond that.

Now produce markdown.

Let's craft.

We need to respect admonitions syntax: four spaces then !!! note "Observacao" then indented text.

We'll include maybe a note: "Observacao: SkillOpt não altera pesos do modelo."

Tip: "Dica: Para experimentar, clone o repositório no GitHub e siga o README."

Warning: "Atencao: Os resultados dependem da qualidade dos trajetos e do split de validação."

Make sure indentation correct.

Now produce final answer with sections: TITULO, RESUMO, FERRAMENTAS, CATEGORIA, TAGS, CORPO.

Make sure each label followed by colon and content on same line? The spec:

TITULO: (titulo descritivo curto, maximo 10 palavras) RESUMO: (2-3 paragrafos explicando o que o post aborda) FERRAMENTAS: (SOMENTE nomes de softwares, frameworks, linguagens, servicos cloud ou APIs reais...) CATEGORIA: (uma so: ...) TAGS: (3 a 5 tags tematicas curtas...) CORPO: (Reescreva o conteudo principal como um artigo markdown bem formatado...)

Thus after each label we put colon then content on same line for title, resumo, ferramentas, categoria, tags. For corpo we put colon then line break then markdown content.

Let's craft.

Title: Microsoft abre o código do SkillOpt para treinar agentes LLM

Resumo: Need 2-3 paragraphs. We'll write two paragraphs.

Paragraph1: Microsoft anunciou a abertura do código-fonte do SkillOpt, um framework que permite treinar habilidades de agentes de linguagem grandes (LLMs) assim como se treina uma rede neural. O SkillOpt considera um arquivo markdown simples como o parâmetro treinável de um agente LLM congelado, aplicando os mesmos conceitos de otimização usados no treinamento de pesos: taxas de aprendizado, portões de validação, tamanhos de lote e agendas de épocas.

Paragraph2: O processo envolve o modelo congelado executando tarefas com a habilidade atual e gerando trajetórias pontuadas; um modelo otimizador analisa falhas em minibatches, propõe edições estruturadas de adição, exclusão ou substituição e as classifica sob um orçamento de edição. Se a habilidade candidata melhorar o desempenho em um conjunto de validação reservado, a edição é aceita; caso contrário, é rejeitada e armazenada para evitar repetições. O resultado final é um único arquivo best_skill.md contendo regras legíveis e compactas, descobertas automaticamente

Ferramentas e Tecnologias

  • [[(only names of real software]]
  • [[languages]]
  • [[cloud services]]
  • [[APIs. No file names]]
  • [[abstract concepts]]
  • [[generic terms. If none]]
  • [[write "Nenhuma")]]

Nota pessoal

https://x.com/i/status/2059993771409015076

Tags

(3-to-5-short-thematic-tags #no-file-names)