The Claude-in-Slack story from yesterday isn't done. Box CEO Aaron Levie is back with a sharper take on what actually makes this different, and it points to a design problem every company building AI agents is about to run into: an AI that works with a whole team needs a fundamentally different architecture than one that works with a single person.
01
The "AI coworker" model needs its own identity, memory, and permissions
Box CEO Aaron Levie dug into the Claude-for-Slack launch and flagged something the initial announcements glossed over. This isn't Claude as a personal assistant that happens to live in Slack. It's Claude as a shared resource that an entire team can invoke simultaneously, which creates real consequences: the agent needs its own persistent memory across conversations, its own permission scope that isn't tied to any single user's credentials, and some way to coordinate when multiple people are pulling on it at once. Levie pointed to agentic coding tools and products like OpenClaw and Hermes as early examples of this pattern, but noted that general-purpose knowledge work raises the stakes considerably. ---
Why it matters: Your IT security team approved AI tools by asking "what can this person's AI access?" The new question is "what can this agent access on behalf of everyone?" Those are different threat models. If your company is deploying Claude in Slack without having answered the second question, you have a gap worth closing before something embarrassing happens.
OpenAI publishes research on how agents are actually changing work
OpenAI released a research paper arguing that AI agents are enabling longer and more complex tasks than previous AI tools, expanding productivity across job roles rather than just automating narrow functions. The paper is light on specifics in the public summary, but the framing matters: OpenAI is positioning agents as a qualitative shift in what AI can do at work, not just a faster version of the same thing. ---
Why it matters: OpenAI publishing research on agent productivity right now is also a sales document aimed at CFOs who need to justify AI spending. Read it with that in mind, but the underlying question it raises is real: if agents can genuinely handle longer, more complex tasks, the productivity accounting your company did six months ago is probably wrong in either direction.
Peter Yang: Claude Design nailed his app screens, then immediately told him to slow down
Peter Yang, a product creator and writer, tested Claude Design by feeding it a mobile app repo and asking it to reproduce the screens. It did it well. Then, after a single follow-up prompt, Claude told him to conserve tokens. He found this funny. It is also a data point. ---
Why it matters: Token limits are the new rate limits, and the tools are starting to enforce them proactively. If you're building on top of Claude and your users are going to have heavy sessions, plan for this now rather than after a confused support ticket from a customer whose AI mid-session.
Swyx highlights a podcast packed with Databricks and agent infrastructure takes
Swyx flagged a podcast episode he called full of "alpha," covering why Databricks beat Snowflake in the data platform wars, why everyone is now building what he calls a "metaharness" for AI agents, the acquisition of MosaicML, and how to hold onto research culture inside a company now valued at $175 billion. The Neon Database acquisition also gets extended treatment. ---
Why it matters: The "metaharness" framing is worth paying attention to. If every serious AI team is building the same layer of orchestration tooling on top of foundation models, whoever sells that layer as a product wins the infrastructure race the same way AWS won the server race. This podcast is apparently where some of that thinking is getting worked out in public.
Quick hit: Thibault Sottiaux says his internal Slack is giving Twitter a run
Thibault Sottiaux posted that his company's internal Slack is "at times at least as fun as Twitter," which earned 643 likes and 141 replies. No further context was provided. Given that Claude just moved into Slack and yesterday's digest was largely about what happens when AI lives where your team already communicates: maybe the Slack-versus-everything-else battle is more interesting than it looked a year ago.