Caywork Platform
Author at Caywork
Read the v23 changelog top to bottom and a pattern emerges: this is not a release of one flagship feature. It is a release about the structure an AI agent platform needs before its features can compound. Three structural bets stand out: everything belongs to a workspace, every external system is one categorized node away, and every piece of retrieved knowledge can be inspected while it is being fetched.
Bet 1: The Workspace Is the Unit of Organization
v23 makes the workspace real. There is a workspace feature and a switcher, agent chats live under workspace-scoped paths, and files upload into the workspace they belong to. Creation is deliberately low-friction: a URL slug derives itself, an advanced toggle keeps the power settings out of the way, and error messages were rewritten to be friendlier.
The feature that ties it together is state persistence across create pages, the dashboard, chat, the knowledge base, and account sections. Interfaces that remember where you were are the difference between a tool you configure once and a tool you reconfigure every session.
Bet 2: Reach Should Be Browsable
The connector story doubled. The MCP connector node library now ships more than fifty integrations, and they arrive in a sidebar organized by category with logos, so discovering that an integration exists is a scrolling problem, not a research problem.
The first-party connectors grew up too. Google Docs, Sheets, Gmail, and LinkedIn nodes are configurable now, and the Google Calendar node received a type update. For anyone building agents that touch real work systems, the practical meaning is simple: the default answer to "can an agent reach this" moved from "let me check" to "probably, check the sidebar."
Bet 3: Retrieved Knowledge Should Be Watchable
The knowledge base changes are the most concrete. Web and website sources are now managed through an interactive UI with live crawl progress: you add a source and watch the crawl as it happens. Website sources are filtered out of the files tab so documents and web sources do not blur together, and the inspect modal streams a full markdown preview of what the crawler captured.
Streaming the preview is the load-bearing detail. Reading exactly what the agent will retrieve turns context from a hope into a check. If a crawl came back thin or a page captured badly, you see it when you add the source, not when an agent cites nothing in front of a customer.
The Agent Builder, Rebuilt
Around those three bets, the builder experience was reworked end to end. Agent creation has new logic with a refreshed detail and flow UI, and the flow sidebar opens by default to hand the canvas more room. Onboarding asks questions and turns the answers into agent settings, powered by a generative question UI node that ships as a template. Sticky notes work on the flow, and a featured agents section gives new users a starting point that is not a blank canvas.
Supporting cast: a Learn resources page and a docs tips and guides page, user avatars in chat with a dynamic expanding layout motion, and a chat sidebar redesigned around agent and session tabs for faster management.
The Polish Layer
The improvements list reads like the first five minutes of using the product, because that is where it acts: restyled single pages under /pages/, error states and empty-state actions for tables, new loading skeleton animations across pages and blog posts, a restyled already-logged-in card that redirects home after five seconds, footer reordering with a status page, blog cover image styling, a visual and narrative presentation for maximum settings, updated email templates, and an updated available model list.
Where to Start
Build a workspace first and let the slug and switcher do their job. Add one website source to the knowledge base and watch the crawl live, then open the inspect modal and read the markdown it caught. Wire one MCP connector node from the sidebar category that matches your stack. If onboarding asks questions, answer them honestly; they become your agent settings.