CrewAI

¡Gracias a todos los que participaron en nuestro primer webinar en español! 🚀 👉 La grabación está disponible en el mismo link que usaron para registrarse. Esto es solo el comienzo — estén atentos a los próximos webinars y eventos. #CrewAI #AgenticSystems #IA #TechCommunity

Every major player in AI just made the same architectural bet. 𝗠𝘂𝗹𝘁𝗶-𝗮𝗴𝗲𝗻𝘁 𝘀𝘆𝘀𝘁𝗲𝗺𝘀. Look at what shipped in the last 12 months: → 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 — sub-agent spawning for subtasks, Agent Teams mode for parallel multi-agent execution across entire codebases → 𝗢𝗽𝗲𝗻𝗖𝗹𝗮𝘄 — 200k GitHub stars, fastest-growing repo in history, built on agent-calling-agent architecture with heartbeat scheduling → 𝗖𝘂𝗿𝘀𝗼𝗿 — runs up to 8 AI agents in parallel via git worktrees, each working on separate parts of your codebase simultaneously → 𝗢𝗽𝗲𝗻𝗔𝗜 𝗖𝗼𝗱𝗲𝘅 — coordinated agent teams across the full software lifecycle, 1M+ developers using it monthly → 𝗠𝗮𝗻𝘂𝘀 — three-layer architecture (Planner/Executor/Verifier), 147T tokens processed, 80M+ virtual machines spun up, acquired by Meta for $2B+ At CrewAI, we've been building multi-agent orchestration since the beginning. It's not a pivot — it's the foundation and a bet we took early days. But here's what most people miss about our approach: 𝗪𝗲 𝘀𝗲𝗽𝗮𝗿𝗮𝘁𝗲𝗱 𝘁𝗵𝗲 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲. 𝗖𝗿𝗲𝘄𝘀 handle the AI reasoning — role-based agents that collaborate, delegate, and reason together. 𝗙𝗹𝗼𝘄𝘀 handle the deterministic backbone — event-driven workflows that control when crews execute, in what order, with what guardrails. The AI reasons. The flow controls. You can mix Python, single LLM calls, and full multi-agent crews in one pipeline — with branching, loops, and real-time state. That separation is why 𝟲𝟬%+ 𝗼𝗳 𝗨.𝗦. 𝗙𝗼𝗿𝘁𝘂𝗻𝗲 𝟱𝟬𝟬 companies build with CrewAI in production. Gartner reported a 𝟭,𝟰𝟰𝟱% 𝘀𝘂𝗿𝗴𝗲 in multi-agent inquiries from Q1 2024 to Q2 2025. The agentic AI market is $7.3B today — projected to hit $𝟭𝟯𝟵𝗕 𝗯𝘆 𝟮𝟬𝟯𝟰. This isn't a trend. It's a 𝗰𝗼𝗻𝘃𝗲𝗿𝗴𝗲𝗻𝗰𝗲. Single-agent systems hit a ceiling. Multi-agent coordination is how you break through it and where the real gains of agentic automations shine. Are you seeing this shift in your own stack or across the industry?

A revolução dos sistemas agênticos no setor financeiro já começou. E não é sobre “automatizar tarefas”. É sobre montar "crews" digitais que: entendem contexto, colaboram entre si, executam processos ponta a ponta, e elevam o nível de eficiência, compliance e experiência do cliente. Foi exatamente isso que deu pra sentir no Safra AI Week. O que mais impressiona não é a tecnologia em si. É a ambição e a velocidade com que o Safra está construindo essa jornada. E, como parceiro, isso anima ainda mais. Porque o impacto é enorme e bem real: do backoffice ao atendimento, da análise de risco à personalização, da produtividade interna à inovação de produtos. Esse é só o começo! Seguimos juntos, aprendendo e construindo. 🚀

Cut token costs by ~90% and reduce latency by restructuring your prompts. Prompt caching is one of the most underused optimizations in agentic systems — and it shows up fast when you build with frameworks like CrewAI. In a typical multi-agent flow (think: planner → researcher → executor), the biggest cost isn’t “reasoning”… it’s re-sending the same giant prefix over and over. The trick: LLM caches work left → right. Keep the prefix stable and every iteration becomes an incremental update instead of a full recompute. How this maps to CrewAI-style flows: ✅ 𝐏𝐮𝐭 𝐬𝐭𝐚𝐛𝐥𝐞 𝐬𝐭𝐮𝐟𝐟 𝐢𝐧 𝐚 𝐟𝐫𝐨𝐳𝐞𝐧 𝐩𝐫𝐞𝐟𝐢𝐱 - tools + role + guardrails + few-shot - agent persona (role, goal, backstories) + operating rules ✅ 𝐔𝐬𝐞 𝐋𝐋𝐌 𝐛𝐞𝐟𝐨𝐫𝐞/𝐚𝐟𝐭𝐞𝐫 𝐡𝐨𝐨𝐤𝐬 𝐭𝐨 𝐞𝐧𝐟𝐨𝐫𝐜𝐞 𝐜𝐚𝐜𝐡𝐞-𝐬𝐚𝐟𝐞 𝐩𝐫𝐨𝐦𝐩𝐭𝐬 - before hook: normalize / dedupe / order messages, strip timestamps, keep tool schemas stable - after hook: append-only logging (don’t rewrite earlier messages), persist structured state outside the prompt ✅ 𝐌𝐚𝐧𝐚𝐠𝐞 𝐦𝐞𝐬𝐬𝐚𝐠𝐞𝐬 𝐰𝐢𝐭𝐡𝐢𝐧 𝐭𝐡𝐞 𝐟𝐥𝐨𝐰 - keep “memory/state” in a store (DB/kv) or within the crewai flow state. - only inject the delta as the last message avoid editing earlier blocks mid-run Result: cheaper, faster agent loops — without switching models provided by the inference. 𝐏𝐫𝐨𝐦𝐩𝐭 𝐚𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 𝐢𝐬 𝐬𝐲𝐬𝐭𝐞𝐦𝐬 𝐚𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞