CrewAI

Join us tonight for Game On: The AI Stack Social — the highest-rated side quest at World's Fair. Catch the World Cup, battle it out on arcade and video games, and hang with the best crowd in AI. Free food, drinks, and swag all included. Co-hosted with Arize AI, Amazon Web Services (AWS), BAND and Yugabyte. 📍 10 min walk from Moscone Center — RSVP for the exact spot 🗓 Tonight | 5–8 PM RSVP link in the comments 👇 #AIEngineerWorldsFair #AgenticAI #CrewAI

Conversational Flows are here, and great work from Lorenze Jay Hernandez on this one. No more giant chatbot loops mixing logic, routing, and prompts. Define your conversation as an event-driven Flow, let CrewAI handle the rest. Available now in v1.15.0. Docs in his post below 👇

Latest CrewAI feature: 𝐂𝐨𝐧𝐯𝐞𝐫𝐬𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐅𝐥𝐨𝐰𝐬. Most conversational AI applications end up mixing business logic, routing, and LLM prompts into one giant chatbot loop. Conversational Flows 𝐬𝐞𝐩𝐚𝐫𝐚𝐭𝐞 𝐭𝐡𝐨𝐬𝐞 𝐜𝐨𝐧𝐜𝐞𝐫𝐧𝐬. You define your conversation as an event-driven Flow while CrewAI handles conversational state, intent routing, and turn management for you. The result is conversational applications that are easier to reason about, test, and maintain while balancing determinism. 𝐀𝐯𝐚𝐢𝐥𝐚𝐛𝐥𝐞 𝐧𝐨𝐰 𝐢𝐧 v1.15.0. You also get an awesome looking TUI to play with! Docs: https://lnkd.in/gegz3tGf

Spent time recently with João (Joe) Moura, founder and CEO of CrewAI and it’s hard to overstate how much of the #multi-#agent infrastructure now running inside #regulated enterprises traces back to what he started building in 2023. CrewAI now powers over a #billion agentic automations a month across #Fortune 500 companies. But the more interesting part of our conversation wasn’t the scale, it was the #architecture underneath it. We spent most of our time on a tension that I think defines the next phase of #agentic #AI in regulated industries: the difference between a fixed pathway and a learning system. Take a #claims #process running multiple specialized agents in parallel, coverage verification, fraud screening, payout calculation, audit. The sequence has to be fixed. Regulators need a predictable, auditable pathway from intake to payout, every time. But the agents executing each step shouldn’t be frozen. #Fraud patterns shift constantly. An agent that never adapts isn’t safer, it’s just predictably wrong in the same ways indefinitely. The real frontier isn’t choosing between rigid pathways and adaptive agents. It’s building the instrumentation that makes #adaptation legible, #versioned behaviour, #explainable deltas, #rollback by design; so #regulators can trust a system that’s still learning. The pathway isn’t the constraint. The black box is. Grateful for the conversation, João, looking forward to more of our work together! Thanks for your time, friendship and partnership ! Love your energy and by the way your Even Realities glasses ! #crewai #deloitte #agenticai #mutiagent #architecture #Ai #innovation Hammad Rafique Jonathan Gray Sulabh Soral Costi Perricos Hayley McKelvey Stacey Winters Charlotte Warburton Richard Eighteen Cillian Bane Patrick Thompson

𝐋𝐨𝐨𝐩 𝐞𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 is the phrase of the month. The idea behind it is not new to us at CrewAI. The people shipping the best agents keep landing on the same conclusion. The leverage is not the prompt. The shape that keeps showing up is three components, stacked on top of each other. 𝐓𝐡𝐞 𝐭𝐫𝐢𝐠𝐠𝐞𝐫. A trigger fires the agent. A new email. A closed deal. A cron tick. 𝐓𝐡𝐞 𝐚𝐠𝐞𝐧𝐭 𝐥𝐨𝐨𝐩. A model calls tools until the job is done. 𝐕𝐞𝐫𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧. Checks the output and sends it back when it falls short. Here is the part worth saying out loud. None of this is theoretical for us. 𝐄𝐯𝐞𝐫𝐲 𝐨𝐧𝐞 𝐨𝐟 𝐭𝐡𝐞𝐬𝐞 𝐚𝐥𝐫𝐞𝐚𝐝𝐲 𝐫𝐮𝐧𝐬 𝐨𝐧 𝐂𝐫𝐞𝐰𝐀𝐈 𝐚𝐧𝐝 𝐞𝐬𝐩𝐞𝐜𝐢𝐚𝐥𝐥𝐲 𝐨𝐧 𝐭𝐡𝐞 𝐀𝐌𝐏 𝐩𝐥𝐚𝐭𝐟𝐨𝐫𝐦 𝐰𝐡𝐞𝐫𝐞 𝐲𝐨𝐮𝐫 𝐚𝐠𝐞𝐧𝐭𝐢𝐜 𝐬𝐲𝐬𝐭𝐞𝐦𝐬 𝐫𝐮𝐧 𝐢𝐧 𝐩𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐨𝐧 𝐞𝐧𝐯𝐢𝐫𝐨𝐧𝐦𝐞𝐧𝐭𝐬. - The agent loop is the core. Agents call tools and keep going until the work is finished. - Events are triggers. Cron and webhooks fire a crew from Gmail, Salesforce, HubSpot, or Slack. Nobody presses go. - Verification is set through expected outputs and structured outputs. The agent fills a schema you defined. The result gets checked, not hoped for. Put it all together, and your agentic systems now operate in a loop with inputs dynamically inserted.

We just launched an experimental feature in CrewAI: 𝐂𝐨𝐧𝐯𝐞𝐫𝐬𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐅𝐥𝐨𝐰𝐬. This is a great example of how reliable agentic systems balance 𝐝𝐞𝐭𝐞𝐫𝐦𝐢𝐧𝐢𝐬𝐦 𝐰𝐢𝐭𝐡 𝐫𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠 𝐚𝐧𝐝 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞. In production systems, not every decision should be left entirely to an LLM. The challenge is determining when to use reasoning and where to enforce structure. Conversational Flows help solve that by abstracting the classification layer, allowing you to focus on defining the routes and behaviors within your orchestration layer. The best part? It’s built entirely on top of 𝐅𝐥𝐨𝐰𝐬. All you have do, is set `conversational=True` Docs: https://lnkd.in/g3T8ys77 Would love to hear what use cases you’re building with it.

At the Databricks Data + AI Summit this week, João (Joe) Moura said something that stuck: The agent loop is roughly 1% of the work. The other 99% is building, securing, and governing it. That 99% is exactly why we're saying this: stop giving your agents database credentials. "Data access" isn't one problem. It's four — semantic queries, governed SQL, registered business logic, and vector retrieval. A single connection string collapses all of them with no audit trail and no governance. We just shipped Databricks as a managed integration in CrewAI AMP — four separate MCP servers, one for each interaction pattern. Agents should work within your existing governance boundaries. Not tunnel around them. Read more → https://lnkd.in/d26yezWK #AgenticAI #CrewAI #Databricks #DataAISummit

Reliability is the new frontier in agentic AI. Our own Lorenze Jay Hernandez just broke this down at Databricks: the best production agents balance determinism and agency. Too much of either and you lose. Great talk, Lorenze. 👇 #AgenticAI #CrewAI #Databricks

Every week in The Context Window we plot what happened in AI on two axes. Slop ↔ Signal: was there actual substance, or noise? Now ↔ Later: act on this quarter, or a signal about where things are headed in 1–3 years? Here’s the matrix for last week 🟢 NOW / SIGNAL OpenAI IPO Filing at $852B-$1T But here's the thing: OpenAI loses $1.22 for every dollar it earns. HSBC projects it needs $207 billion more by 2030 just to stay operational. The Pope just declared war on AI 1.4 billion Catholics now have a manifesto against Silicon Valley. "Magnifica Humanitas" was released May 25 with Christopher Olah (Anthropic co-founder) at the podium. IBM's Watson Orchestrate just shipped multi-agent orchestration for the enterprise. "Many have invested heavily in AI, but few believe it's paying off." IBM's solution is to stop deploying agents in isolation and start wiring them together. That's where the ROI actually lives. 🔴 NOW / SLOP Sam Altman says he was "pretty wrong" about the AI jobs apocalypse timing: 8 days after filing for IPO. Twelve months earlier, he warned that entry-level white-collar roles were "at serious risk." Well as per data, 115,000+ tech layoffs have already made in 2026. Kim Kardashian posted an AI recreation of her late father and spawned a viral deepfake surge. Deepfakes went from "YouTube prank" to "emotional manipulation at scale." The line between fandom and exploitation is going to collapse soon. Google's experimenting with celebrity-inspired AI chatbots without asking permission. No licensing deal, or consent disclosure. This won't last legally, but it signals where the math is heading: if you can generate the persona, why pay for the person? 🟢 LATER / SIGNAL AI Agents market is growing 10x by 2030: $5.25B (2024) → $52.62B (2030). The fragmentation is obvious in the use cases: coding agents, customer-service agents, healthcare agents, legal agents, security agents, procurement agents, finance agents. AI Drug Discovery entering the mega-round phase Isomorphic Labs announces $2.1 Billion funding to power its AI drug design engine, scale its business globally and progress its drug candidate pipeline. 🔴 LATER / SLOP Telegram's launching AI assistant bots, another messaging platform betting on persistent agents. Messaging platforms keep launching AI "layers" every 18 months and sunsetting them when usage flatlines. Wikipedia editors are threatening a strike over AI-driven layoffs,  regulatory theater in real-time. The actual impact on the AI industry: zero. The symbolic impact on regulation: everything. Apple & Meta are "discussing" integrating Meta's Llama into Apple Intelligence no deal yet, just garden-variety tech gossip. Apple already has a deal with OpenAI. This rumor will either resolve in a licensing agreement or evaporate. 🔗 full edition in the comments.

Chicago. June 24. Let's build. 🏙️ Walk away with a working enrichment agent — ready to deploy from day one. We're bringing a hands-on CrewAI workshop to Chicago — the fastest way to go from idea to working agent. In one session, you'll go from zero to a fully functional agent you can take back to your team. 📍 Convene, Chicago 🗓 June 24 🎟 Grab your spot