Slack launches Slack Code for vibe-coding with AI agents
Three threads from today's feed: collaboration tools are bending toward agents, a founding figure of reinforcement learning is questioning the field's favorite shortcut, and the cybersecurity job market is splitting along an experience line.
Slack Code turns chat channels into shared workspaces for coding agents. The new feature gives teams dedicated, project-specific channels where anyone can tag in a coding agent — Anthropic's Claude or Cognition's Devin — to build a feature, fix a bug, or update a page. The agent spins up a channel, writes the code in the open, and the whole team gets visibility: they can read the conversation, review the diffs, preview the output in a live HTML view, and approve the work before it ships. When the task is done, the channel archives itself and leaves an audit log.
This is Slack's bet that "vibe coding" stops being a solo hobby and becomes a team sport. The pitch is deliberately framed around agents as teammates — you don't paste code into a thread, the agent lives in the channel and everyone audits it together. It's the same "agents as coworkers" framing SpaceXAI is pushing with its Grok Bot, and it tells you where the incumbents think the interface is heading: not a separate IDE, but the place you already talk. The real question is whether teams trust an agent enough to rubber-stamp its diffs, or treat the channel as another inbox to babysit.
Richard Sutton calls synthetic data "a big mistake" for scaling AI. The Turing Award winner and author of "The Bitter Lesson" says leaning on machine-generated training data won't break through the limits of today's language models. His argument, built on the "Big World Hypothesis," is that the real world is infinitely more complex than anything we can simulate, so any synthetic dataset is microscopic by comparison — and deciding which synthetic data is even good requires human experts, which just reintroduces the bottleneck models were supposed to escape.
Sutton's fix is to take humans out of the loop and let agents learn continually from their own experience, building their own world models and correcting them instead of relying on frozen weights. He calls LLMs "like 20% or a quarter of intelligence" — an amazing breakthrough, but a partial one. It's a pointed counterweight to the industry's recent enthusiasm for training on model-generated text; we broke down the mechanics in AI 101 — What is synthetic data?. Whether you buy his framing or not, it's a reminder that the "just generate more data" era may have a harder ceiling than the labs assume.
AI skills in cybersecurity jobs doubled across the G7 as junior hiring stalls. According to the AI Workforce Consortium's first spotlight report, the share of cybersecurity postings asking for AI skills rose to 28.5%, up from 14.2% a year earlier, while overall demand across Group of Seven economies grew 9.5% in the six months through March. The catch is who the jobs are for: senior-titled roles grew 65% over that window, while junior-titled roles grew just 5.9% — a gap the consortium calls the "experience paradox," where employers want senior skills for nominally entry-level work.
The shift tracks a real change in the job itself. The tier-one security-operations-center analyst, once manual triage, is now closer to supervising agents and checking their work — as Cisco's Omar Santos put it in a consortium webinar, "everybody is an orchestrator now." For workers, the message is blunt: hands-on experience with AI agents is the top gap hiring managers cite in entry-level candidates, ahead of technical depth. The people who learn to direct the agents, not compete with them, are the ones the market is paying a 14.9% salary premium to hire.
Is "agents as teammates" the interface we actually want, or just the one incumbents can ship first? Tell us in the comments.
