TL;DR
Altman said smarter models still beat cheaper ones. Anthropic shipped Managed Agents for long-running Claude work, while Masad and Shipper pushed the same thesis: agents are turning prompting into orchestration and sitting beside every app.
Crabbox gets remote Linux runs and replay
Crabbox 0.3.0 is out, and it’s turning dirty worktrees into remote Linux runs with GitHub browser login, live run replay, durable events, AWS image creation, and Cloudflare Access. He also says the package got leaner by moving almost everything into extensions, which should help with the npm dependency/slowness complaints.
AI doesn’t kill jobs — it unlocks pent-up demand
He argues AI won’t replace engineers so much as make them 2x–5x more capable, which pushes companies to hire more and tackle projects they previously couldn’t afford. The bigger point: once every bank, manufacturer, retailer, and SMB gets the same model leverage tech companies have, a lot of “unmet needs” suddenly become worth building for — across engineering, marketing, legal, finance, and design.
Oakland taxes high, services still lousy
He says Oakland is charging more than comparable cities while delivering some of the worst services, with 44% of Measure E revenue already earmarked for union raises. His broader point: dysfunction gets dressed up in moral language, and anyone asking where the money went gets treated like the problem.
AI agents are becoming file janitors
He’s using Codex and Claude Code to clean up local files and Google Drive like a digital Marie Kondo — but only after asking for a plan first, since these are semi-dangerous ops. The takeaway: with enough permissions, AI is already useful for boring admin work that most people keep putting off.
Enterprise sales and token resale win, not vibes
They said Vibe-kanban’s shutdown is a useful postmortem: even with 30k MAU, the business missed the two things that seem to matter most right now — enterprise sales and token resale. The bigger takeaway is the software-engineering retrospective from 2021–2025: a lot of “cool” products still die when the monetization model doesn’t.
VCs are optimizing for downside, not returns
He says too many funds are doing deals for the safety of a token-factory bailout instead of actually swinging for 5–10x outcomes. The bigger gripe: incentives have drifted toward AUM-maxxing and “deploy baby deploy,” which he thinks is a short-sighted way to waste time and capital.
Parallel agents turn prompting into orchestration
He says “prompt” now means something bigger, but the core game hasn’t changed: you’re still steering systems, just with more moving parts. His other note — 10 projects, 10 parallel agents each — points to the Replit CEO’s view of AI work shifting from single-shot prompting to managing swarms of agents.
Continuous agents will sit beside every app
He says this is the default work pattern for the next decade: an agent running continuously on the left, with the app you and the agent use on the right. It’s a clean codex-native workflow thesis from the Every CEO — and he’s already pointing people to a tool called Proof to try it.
Scaling Managed Agents: Decoupling the brain from the hands
Anthropic launches Managed Agents for long-running Claude work
Lead: Anthropic introduced Managed Agents, a hosted Claude Platform service that decouples the agent “brain” from its “hands” and session state so long-horizon work can run through stable interfaces even as models and harnesses change.
Numbers:
- p50 time-to-first-token dropped roughly 60%.
- p95 time-to-first-token dropped over 90%.
- The system exposes a small set of interfaces: `execute(name, input) -> string`, `provision({resources})`, `wake(sessionId)`, `getSession(id)`, and `emitEvent(id, event)`.
So What: For builders, the key shift is architectural: Claude’s harness no longer lives inside the sandbox/container, credentials stay out of untrusted code, and session state becomes durable and inspectable outside the context window. That makes failures easier to recover from, supports VPC and custom-tool integrations, and lets one brain talk to many hands without coupling everything to a single container. Anthropic’s framing is explicit: “The harness leaves the container,” and the goal is a “meta-harness” that can outlast any specific implementation. In practice, this means you can plug in Claude Code, task-specific harnesses, MCP tools, or future sandboxes without redesigning the whole stack.
Greg Brockman says attention, not code, is the real constraint
The Takeaway: AI is making execution cheap, so the scarce resource is now human judgment.
- OpenAI’s real bottleneck isn’t ideas or demand — it’s compute, and Brockman says they still want more of it.
- The big shift isn’t just smarter models; it’s context. If the AI doesn’t know what’s happening, you’re wasting its power.
- The hardest problem ahead is governance: deciding what to auto-approve, what to escalate, and what still needs a human signature.
Greg Brockman, OpenAI cofounder and president, frames the company less like a lab and more like a machine for turning compute into intelligence. His line is blunt: “We buy, rent, build, compute, and we resell it at a margin.” That model only works because demand keeps outrunning supply — and because the models keep getting better as more compute is poured in.
But the sharper insight is about how work changes around the models. Brockman argues that the breakthrough isn’t just capability, it’s context: AI is becoming useful when it can see your meetings, your files, your history, and your workflows. Otherwise, you’re stuck doing the dumbest part yourself — explaining the situation to the machine. That’s why he thinks the next wave is about harnesses, memory, and systems that let AI act inside real work, not just answer prompts.
He’s also unusually clear that speed needs guardrails. OpenAI still wants a human accountable for merged code, and Brockman says the company is building security, observability, and escalation systems because “human attention is going to be this incredibly scarce resource.” The future, in his view, isn’t humans replaced — it’s humans promoted to judge what matters while agents handle the rest.
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.