TL;DR
Claude doubled token limits and added everyday-life connectors, while Anthropic’s Managed Agents got cross-session memory and a stronger runtime feel. Rauch made Vercel firewall protection free, and Levie said agents fail from messy context, not missing context.
Claude says you can now create more with Claude Design, and it’s bumped token limits across every plan. That’s the kind of quiet product upgrade power users notice fast — more room for bigger prompts, longer docs, and fewer annoying cutoffs.
He’s pushing a simple fix for agentic coding: keep a running `implementation-notes.html` so the model can log ambiguities, deviations, tradeoffs, and open questions as it works. The point is to let Claude make decisions when specs get fuzzy, without leaving humans out of the loop.
Vercel is now eating the compute and network costs for firewall mitigations, including custom rules — not just DDoS and system-level defenses. He also called out the product’s ~300ms global propagation, a big deal if you’ve ever waited minutes for CDN/WAF changes to land.
He says the real bottleneck in enterprise agents is constrained context: too much conflicting info and they hallucinate, too little and they’re useless. That makes most AI strategy problems really data strategy problems — which is why structured and unstructured data hygiene matters so much for big companies, and why startups can win by designing for it from day one.
He says GBrain is moving fast day to day, and the latest push bundled 28 bug fixes from 22 community PRs and 14 issues. He also published the full eval report and fixtures as open source, inviting other memory systems to benchmark against them and get listed alongside.
Anthropic is treating the model and its harness as a coupled system, so the same Claude can behave differently across Claude, Cowork, and Claude Code. The bigger takeaway: they’re using Claude to cluster user pain, generate synthetic evals, and even “dream” over memories when idle — while some researchers are seriously thinking about consciousness as agents get more autonomous.
He says Everlane’s real problem was the pandemic: the whole wear-to-work category vanished overnight, and the financing needed to survive made the endgame uglier. The bigger take is that brands can look dead for years, then come back hard — he points to Ray-Ban, J.Crew, Madewell, and Abercrombie as proof.
Anthropic fixes three Claude Code regressions and resets limits
Lead: Anthropic says recent quality complaints about Claude Code came from three separate product changes—not a model or API degradation—and all have been fixed as of April 20 (v2.1.116).
Numbers:
So What: Builders should treat this as a reminder that product-layer tweaks can look like model regressions, especially when they hit different user slices on different timelines. Anthropic says the API and inference layer were unaffected, but the combination of effort defaults, cache handling, and prompt edits made Claude Code feel “less intelligent,” forgetful, and more token-hungry. The company is tightening prompt-change controls, expanding per-model evals, improving internal code review, and adding soak periods plus gradual rollouts for any change that could trade off against intelligence. As Anthropic put it, “This isn’t the experience users should expect from Claude Code.”
Claude adds everyday-life connectors for travel, shopping, and more
Lead: Claude is expanding its connector ecosystem beyond work tools to include everyday apps like AllTrails, Instacart, Audible, Tripadvisor, TurboTax, Uber, and more, so users can act on personal tasks directly inside chat.
Numbers:
So What: This turns Claude into a more useful action layer for daily life: it can suggest the right app in context, combine multiple services in one thread, and keep users in control before anything is booked or purchased. Anthropic says, “Claude suggests connectors and makes recommendations. But you stay in control of its actions.” The practical takeaway for builders is that connectors are now a distribution surface for consumer apps inside Claude—if your product would be useful in chat, you can submit it to the directory. Users can install connectors quickly, and once connected, they’re available in every conversation without leaving the thread.
Claude Managed Agents get built-in cross-session memory
Lead: Claude Managed Agents now include a public beta memory layer that lets agents learn from every session while keeping developer control through file-based, API-manageable storage.
Numbers:
So What: This turns memory from a custom retrieval problem into a built-in production feature for long-running agents. Because memories are stored as files, teams can export, roll back, redact, and manage them programmatically, while sharing stores across agents with different access scopes. Claude also mounts memory directly onto the filesystem, so agents can use the same bash and code execution tools they already rely on. As Anthropic puts it, “Memory is built for enterprise deployments, with scoped permissions, audit logs, and full programmatic control.” For builders, the practical move is to use Managed Agents when you want agents to retain context across sessions without standing up your own memory infrastructure.
Anthropic’s platform is becoming the agent runtime, not just an API
The Takeaway: The real platform play is to hide harness and infrastructure so agents can ship fast and scale cleanly.
Angela, head of product for Claude’s platform at Anthropic, and Caitlin, head of engineering, describe a platform that’s evolving from completions to stateful, autonomous systems. Their view is blunt: customers don’t really want to rebuild the boring parts. They want outcomes. As Angela put it, the platform should be “the set of primitives and infrastructure that enables you to basically get the outcome as fast as possible, with actually as little of work as possible.”
That’s why Claude Managed Agents bundles the pieces most teams end up reinventing anyway: messages API, code execution, web search, memory, sandboxing, and now vaults for credentials. The surprising insight is that the bottleneck has shifted. People think harness engineering is the hard part, but in practice they hit an infrastructure wall when they try to productionize. A Mac mini and a Python loop can get you a demo; they won’t get you a reliable product.
The bigger bet is that model and platform will converge. Instead of a generic wrapper that swaps models underneath, Anthropic expects the harness and model to become increasingly paired — because path dependence matters, and the “small” choices about file systems or tool use end up shaping what the model becomes good at. The endgame: agents that are easy to deploy, maybe even one-click into Slack, with the platform doing the unglamorous work behind the scenes.
Peter Steinberger called AI code review the killer dev workflow, while Thariq said Claude Code turned files into a workbench. Cursor bet on specialized models plus brutal RL infra, and Anthropic detailed how it contained Claude across products.
Steinberger said AI agents need lean skills, not bloated prose; Levie argued AI won’t kill jobs, it’ll raise the bar. Yang said Codex tested itself while Claude still owned frontend, and Garry Tan pushed hard evals as the way agents got better fast.
Levie warned AI demos fool CEOs; the hard work starts after. Tan said high agency plus taste beat everything, while Yang urged builders to build the system that builds the MVP first. LeCun kept the pressure on: LLMs help, but they’re not the road to real intelligence.
Levie said AI kills tasks, not jobs, so roles just expand. Masad said Replit’s AI stack can ship MVPs in a weekend, while Tan showed Qwen fine-tunes in hours and personal AI got real. No Priors said Cerebras won by betting on speed before the market cared.
Levie said AI will surface more bugs, making humans the bottleneck. Steinberger argued GitHub’s native PR limits kill bot policing, while Tan pushed the “ship the 60%” rule and Swyx said AI coding should harden codebases, not just move faster.