Jerome Cukier

Buried in the Claude Haiku 4.5 system card was a really interesting nugget that most people missed, but it could have a huge impact for anyone building with LLMs “For Claude Haiku 4.5, we trained the model to be explicitly context-aware, with precise information about how much of the context window has been used. This has two effects: the model learns when and how to wrap up its answer when the limit is approaching, and the model learns to continue reasoning more persistently when the limit is further away." For years, context awareness and management have been handled by developers and it's one of the hardest parts of building agents. Teams built systems to count tokens, monitor context usage, and trigger downstream functions like summarization or condensation. With Claude 4.5 Haiku, context awareness is now trained into the model itself. It reminds me of what happened with Chain of Thought. Something that started as a prompt engineering hack but eventually became part of model training (reasoning models). There’s still plenty of work to do around context management, but if this is the direction things are heading, the next major leap in model capability may already be taking shape: models that can manage their own context intelligently. S/O to Simon Willison who originally highlighted this on his blog (linked below)

22

3 Comments

🚨 Claude just took a very nice step forward. Anthropic launched Claude Integrations, letting Claude connect directly to the tools your teams already use—like Asana, Google Workspace , Atlassian's Jira and Confluence and more. 🧠 They’ve also expanded Claude’s Research Mode. It now pulls from the web, your internal docs, and your connected tools—combining information into multi-step, well-sourced responses. Less like a chatbot. More like a smart, embedded teammate. 🧩 Why this matters for enterprise AI strategy: • In my AI readiness work, I see the same pattern again and again: most companies don’t struggle with access to AI—they struggle with integration. • This update reflects a bigger shift: from prompt-based AI to workflow-integrated AI. • Claude is stepping into the flow of work—pulling context, responding to the real state of the business, and supporting action, not just answers. • It’s more than a product drop. It’s a signal of where AI is headed—and what orgs need to get ready for. 📌 Key takeaway: If your AI strategy isn’t planning for integration with your systems and workflows, you’re underestimating the opportunity. 👇 How are you thinking about embedding AI into the way your teams actually work? #ClaudeAI #AIReadiness #EnterpriseAI #AIOrchestration #LeadershipInTheLoop #AIProductivity #AIworkflows #AIEnablement #Jira #Confluence

7

After a few days to reflect, I wanted to share some thoughts on an amazing week of back-to-back conferences in NYC – AI Native DevCon and AI Engineer Code Summit. (It was especially exciting given the flood of big releases – Gemini 3, Nano Banana, Antigravity, etc. etc.) The two events were great complements to one another: AIND was more community and open-source focused, and AIE was more about the leading AI coding products and platforms were headed and what businesses big and small were doing to evolve to take advantage of them. However, a few major themes spanned both: DON'T OUTSOURCE THE THINKING Online conversation might be dominated by “one-shotting” prompts, but last week the vibe was very… anti-vibe(coding). swyx ​set the tone calling for “no more AI slop”, and the theme persisted with many talks focusing on how individuals and teams can better set up their coding tools and agents for success through careful planning and context management. Dexter Horthy from HumanLayer put it most succinctly: “don’t outsource the thinking”. CONTEXT IS ALL YOU NEED How do we avoid poor quality, tech debt, and slop? Despite huge advances in foundation models, better LLMs are not the answer. To quote Steve Yegge: “the models are smart enough”. The answer? Better planning, process, and steering, which is ultimately context: how do you make sure the LLM has access to the right information at the right time at every step? A huge part of the context that coding agents need? Specs, e.g. explicit instructions around expected product behavior, or how to interface with particular versions of libraries or services. AIND (hosted by Tessl) had plenty of killer content about spec-driven development. EVERYONE IS CONCERNED ABOUT ROI Both conferences had insights from real-world data as well as war stories from the field, all with the same lesson: if you want to get the most out of AI coding tools and agents, you need big investments. And not just compute, but in infrastructure (to provide better context) and in organizational and culture changes (training, new roles, new team structures). Martin Harrysson and Natasha M. from McKinsey gave a great talk summarizing what they’re seeing as the organizational and technical investments necessary to become AI native – and what the gains are. Why does this matter? Not only because these huge investments can potentially unlock big dividends, but as Yegor Denisov-Blanch from Stanford showed in his talk, “the rich get richer”: teams that are mastering AI-native development are accelerating and widening the gap. Those were the big takeaways! The quality of both events was excellent and all the content is available online (or will be soon), so definitely check it out for yourself. And if you’re interested in some hard data about what productivity gains look like in the real world check out my Jellyfish AI Engineer Code Summit talk: https://lnkd.in/em3AaFTh

8