AI Builders Digest

Claude Cowork folds into chat, and Docs and Slides ship in beta

Cowork and chat are now one Claude, rolling out to Pro and Max over the next few weeks with Team and Free to follow and Enterprise admins getting 30 days notice. Claude Docs and Claude Slides are new, and Claude Design now runs inside conversations, so a report and the deck that matches it come out of the same thread. You can edit directly, present from Claude, or export to PowerPoint or PDF, and everything lands at one shareable link. The reason given: people used both products and the frustrating part was deciding where a task belonged, with nothing carrying between them.

• #products
• #agents

Claude now routes between quick answer and deep agentic work itself

The merge rests on a bet that the model is good enough to decide for you. Based on your prompt, Claude picks whether to fire back a quick answer or go do longer agentic work, and what output format fits, rather than making you pick a product first. Users kept saying they did not want to choose which Claude to open. You can still stop it, redirect it, or dial its effort up and down mid-task.

• #products
• #agents

The pitch for merging: one Claude that carries context everywhere

Claude Code proved developers could hand over a feature and come back to shipped code, and Cowork proved knowledge workers could hand over a brief and come back to finished files. Merging chat and Cowork is the step toward one Claude that carries context across everything you are working on. Docs, Slides, and Design are in every conversation now, so asking for a presentation returns something you can open, edit, and export as PowerPoint or PDF with no separate tool to navigate to.

• #products
• #agents

Vercel's fx command reviewer hits 18x faster p95 on a smaller model

fx runs in auto mode by default with a safety reviewer analyzing every command before it executes, and that reviewer currently runs on GPT Luna. Swapping to Jev made it up to 18x faster at p95 and more accurate at the same time, which is the rare case where you do not trade quality for latency. It is heading to Vercel AI Gateway and will likely become the new default.

• #agents
• #products

Cheap fast inference unlocks the gates inside enterprise workflows

Entire universes of AI innovation exist that were not on most people's radar, and processing information very quickly at very low cost with high capability changes what is economical in the enterprise. The targets are the gates in a process rather than the headline work: data classification, routing decisions inside a workflow, domain-specific judgment calls, quick safety or security calls. Those are exactly the chokepoints that make agentic workflows viable at scale.

• #agents
• #products

Bash is no longer all you need for reliable tool calling

A reversal from last year's thinking: if the goal is reliable tool calling, wrapping everything in bash is the wrong layer of indirection. Give Claude tools shaped the way you actually want them instead. For data storage, ask whether you really want a filesystem or an API to a database. Sandboxes plus bash still earn their keep for work involving code generation and execution, and Claude Managed Agents strikes the balance by making the sandbox optional and spinnable independently of the agent loop, which held up when porting an old bash-heavy project over.

• #agents
• #open-source

Astra positioned as fast, frontier, efficient, and for everyone

Astra is out, framed on four axes at once: fast, frontier, efficient, and available to everyone. The accompanying note that sometimes physics cannot be cheated reads as a nod to the tradeoffs that do not go away no matter how the model is tuned.

• #products

Eight AI skills now produce an entire podcast, taste still applied by hand

A full podcast pipeline runs on custom skills: /podcast-prep researches the guest and drafts an interview guide, /podcast-edit reads the raw transcript and picks intro reel quotes and cut points together with the host, and /podcast-production orchestrates five more skills to turn one episode into six assets. The contrarian bit is on skills themselves. Some people argue the latest models make them redundant, but they are what gets AI to follow specific editing and browser-use instructions rather than generic ones.

• #agents
• #products

Safety and security are product features, not externally imposed guardrails

Treating safety and security as constraints handed down from outside gets the framing backwards. They are features of the AI product and the model itself, built in as capability rather than bolted on as compliance.

• #policy
• #products

Claude's prose is driving at least one builder away

Reaching for Claude less lately because the way it speaks has become insufferable. The complaint is specific: it reads as constantly showcasing how smart and sophisticated it is rather than actually communicating the point.

• #products
• #evals

A months-in-the-making home agent works because two people share it

A Nous Research based home agent has finally gotten good after months of tinkering, and it is the first group bot in the house, used daily by both partners, which surfaces interesting behavior mostly when it fails. Attempts to recreate it with Grok Bot and similar could not reach the same granularity and control: restricting it to certain emails only, converting inline attachments into structured data, working through logged-in browser sessions.

• #agents
• #products

Any harness, same personality, full memory

The appeal of portable memory put plainly: pick whatever harness you want and get the same personality with the full memory of your personal AGI intact. The client stops mattering once the model carries the continuity.

• #agents
• #products

Decks, docs, and designs are editable in the thread and exportable

Draft a one-pager in Claude Docs, turn it into a deck with Claude Slides, and mock up a matching visual in Claude Design without leaving the conversation. You can fix a line in the doc yourself, leave a comment for Claude on a slide, or select an element and move it, then present the deck straight from Claude, export the doc in your usual format, or share it all at one link. Design, Slides, and Docs are in beta on all paid plans, and anything already in Cowork carries over: chats, projects, artifacts, connectors, and skills.

• #products