Yesterday we wrote about Claude moving into Slack as a shared team agent. Today, Aaron Levie explains why that framing actually undersells what's happening, and OpenAI drops a research paper trying to quantify it. The "AI as coworker" thesis is hardening from metaphor into architecture.
01
Claude in Slack isn't a chatbot integration. It's a new kind of coworker.
Box CEO Aaron Levie followed up on yesterday's Claude-in-Slack launch with a thread worth reading carefully. His point: the significant thing isn't that Claude is in Slack. It's that Claude exists as a shared entity any team member can address, with its own identity, its own memory, and its own context across conversations. It's not you having a 1:1 with a tool. It's more like a new hire who sits in every channel simultaneously. Levie notes that agentic coding tools like OpenClaw and Hermes are already building around this pattern, and that general-purpose knowledge work is next. ---
Why it matters: If your company is planning an "AI tool" rollout, you may be solving the wrong problem. The question isn't which software seat to buy. It's how you onboard, direct, and govern an entity that your whole team shares and that accumulates context over time. That's closer to an HR question than an IT question.
OpenAI publishes research on how agents are actually changing work
OpenAI released a research paper on how AI agents affect productivity across different roles, with a focus on longer, more complex tasks rather than the simple single-turn prompts most benchmarks measure. ---
Why it matters: Most "AI productivity" claims are based on narrow tasks someone chose because AI performs well on them. If OpenAI's data covers the messy, multi-step work that actually fills people's days, that's a more honest test. Watch for which job categories show the biggest gains, because that's where the first real restructuring happens.
Swyx flags a podcast with actual answers about the AI infrastructure wars
Swyx highlighted a new episode covering some surprisingly direct takes: why Databricks beat Snowflake, why every company is now building what he calls a "metaharness" for managing AI agents, and why Neon's database architecture made sense for the agent era. The episode also covers what happened to MosaicML after the Databricks acquisition and how to preserve research culture inside a company now valued at $175 billion. ---
Why it matters: The "metaharness" observation is the one to track. As companies run dozens of AI agents across different tasks, they need a layer that coordinates, monitors, and corrects them. Whoever builds that coordination layer well is sitting on infrastructure that every enterprise will eventually need. The database wars of the 2010s may be repeating at the agent orchestration level.
Peter Yang: Claude Design reproduces mobile app screens from a repo, then immediately asks you to slow down
Peter Yang shared a quick experience with Claude Design: he fed it a mobile app repository and it reproduced the screens accurately. One prompt in, Claude started suggesting he save tokens. The screenshot he attached makes the moment funnier than it sounds. ---
Why it matters: Claude nudging users about token consumption after a single prompt is the AI equivalent of a contractor finishing one wall and asking if you really need the other three. It's a real constraint worth knowing about before you plan a design sprint around it.