Follow builders, not influencers.
TL;DR
Rauch said best product wins and AI’s just the lever. Thibault Sottiaux called GPT-5.5 the best model yet, while Claude’s Managed Agents added dreaming, outcomes, and orchestration. Peter Steinberger noted agent prompts now ran for hours, not minutes.
He says the goal isn’t to maximize AI usage — it’s to ship the best product, whether that means lots of AI, a little, or none at all. He also pointed to per-API-key spend caps in Vercel’s AI Gateway, a practical control for teams trying to keep AI costs from running wild.
He says OpenAI’s GPT-5.x naming isn’t just cosmetic: each bump should mean real gains in capability and token efficiency, which also makes things faster. He calls GPT-5.5 their best model yet and frames the whole strategy as a simple one they want to keep repeating.
Cursor’s auto-review doesn’t just block or allow commands — it explains what they do and why they’re risky, which makes it way easier for new coders to learn by doing. That’s a small UX tweak with a big payoff: less fear, more forward motion.
With GPT-5.5, /goal, autoreview, and crabbox, prompts that used to take 30–60 minutes are stretching into 4–10 hour tasks — and he says confidence in the output is way higher. His take: yielding agents is a skill, and code review gets much better once you push the model past the first “looks fine” answer.
He says enterprise AI is mostly showing up as reinvestment, not layoffs: savings from automation are getting pushed into sales, marketing, engineering, and other underbuilt areas. The bigger point from the Box CEO is that AI expands what companies can do, so the winners will be the ones that use the efficiency to serve customers better — not just trim headcount.
He says Codex is now chewing through 38B tokens, with a 56-hour longest task and a 41-day streak. The vibe: AI coding isn’t just autocomplete anymore — it’s becoming a persistent worker that can stay on a problem for days.
Claude Managed Agents adds dreaming, outcomes, and orchestration
Lead: Claude is launching dreaming in Managed Agents as a research preview, while also shipping outcomes, multiagent orchestration, and webhooks to help agents self-improve, verify their work, and split complex jobs across specialists.
Numbers:
So What: Dreaming adds a scheduled memory-refinement loop that reviews past sessions, extracts patterns, and can either auto-update memory or route changes for review. Outcomes gives builders a rubric-based grader in a separate context window, so agents can self-correct without being biased by their own reasoning. Multiagent orchestration lets a lead agent delegate work to parallel specialists with their own models, prompts, and tools, with persistent events and full traceability in the Claude Console. As the post puts it, “Together, these updates make agents more capable at handling complex tasks with minimal steering.” For teams building long-running, high-stakes workflows, the practical move is to use dreaming for cross-session learning, outcomes for quality control, and orchestration for parallel execution.
Google bets the next leap is world models, memory, and self-directed agents
The Takeaway: Google’s frontier bet is simple: make models that understand the world, remember over time, and build their own scaffolding.
Key Insights
The Story
Oriol Vinyals, co-lead of Gemini at Google, frames the frontier as a shift from clever demos to systems that actually accumulate understanding. His core argument is that language models already benefited from the internet’s giant text corpus, but vision and video still haven’t had their equivalent “aha” moment. Google’s Omni is his proof of progress: it can take in images and video, generate video, and edit it through language, but he says the field still hasn’t unlocked the deeper transfer from raw visual data into compact concepts.
He’s especially interested in world models as more than representation learning. In his words, the goal is to “simulate” the world well enough that models can predict before acting. That’s why robotics keeps coming up: not because today’s models can do precise motor control, but because they may soon help with planning, scenario generation, and gross-level decision-making.
On agents, Vinyals is blunt that the future probably won’t be a pile of brittle hand-coded scaffolds. Instead, “the model itself could write [the system] on the fly.” He sees memory the same way: working memory is already strong, but durable learning will likely live in file-system-style external storage, not constantly rewritten weights. That’s the practical path to continual learning—and maybe the next real paradigm shift.
Boris Cherny said agentic workflows beat speedups by a mile. Garry Tan argued money amplifies demand, not creates it. Aaron Levie called a $500M custom build a bullish ad for software, while No Priors warned agents need guardrails before they become enterprise liabilities.
Matt Turck said enterprise AI had moved from chatbots to messy, expensive agents, and the bill was coming due. He framed the shift as real progress with real pain: more autonomy, more complexity, more cost.
Every argued AI lowered the floor but raised demand for experts. Anthropic fixed three Claude Code regressions, reset limits, and shipped Managed Agents to split agent brains from tools. Claude also added self-hosted sandboxes and private MCP tunnels.
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.