Claude Cowork: 5 things came along that you never saw
5 changes from the last 2 weeks share a single pattern. The agent’s work now carries people, rules, and promises that were never part of your session.
First, a note on the name
On September 16, Anthropic merged Claude chat and Cowork into a single window. You no longer choose an entry point. You type, and Claude decides whether to answer in conversation or to take on a task and execute it across your files and tools.
This library keeps the name Cowork. It now refers to the moment Claude stops talking and starts doing, the execution mode inside the single product. The 30 principles describe what happens to a person when work is delegated to an agent, and delegation still happens. Principle 05, the Right Tool, still separates thinking from executing. The interface no longer marks that line for you, so you mark it yourself: decide that the idea is clear enough before you let it run.
Now the 5 changes.
The house has other people in it
Google opened its Google Home MCP server to early access on September 25, and names Claude Cowork as an example client. An agent can now list rooms, read device state, run commands, and pull event history from your home.
Every earlier delegation touched space you owned alone. A house is shared. When the agent switches off the living room lights, the person reading on the sofa never opened a session. Principle 02, the Delegation Boundary, gains a population as well as a perimeter.
Your promises carry a version number
On August 25, Claude’s chat memory was connected to Cowork, with no option to keep the 2 separate. Anyone who told a client that each session starts fresh made a statement about a tool that has since changed.
A promise about your own conduct stays true as long as you keep it. A promise about a tool’s behavior expires with the next release. Principle 14 says the output is yours even when you did not control every step. So are the commitments.
Expertise now arrives by installation
Veeva announced a Study Builder Agent that configures clinical studies from a protocol, delivered as a Claude Cowork plugin from a repository Veeva manages. Tax, legal, and accounting vendors are building the same kind of product.
The plugin makes hundreds of decisions before you see the output, and it updates on the vendor’s schedule. Principle 23, Expertise as Evaluation, now means reading the plugin’s decisions and running it against a case you know by heart.
Your document has a third author
On September 22, Typeface made its brand-rule system available to enterprise customers, including inside Claude and ChatGPT. Brand guidelines now reach the agent as context, next to your own instructions.
Get Mihailo Zoin’s stories in your inbox
Join Medium for free to get updates from this writer.
When a cheerful line about excellence appears in your apology email, it may come from a rule written for campaigns. Principle 11, Blurred Authorship, used to describe 2 parties. It now describes 3, and the third is invisible unless you ask the agent to attribute each sentence.
The slow report taught you the numbers
Anthropic’s marketing operations team cut its weekly report from up to 2 days to up to 2 hours. Their version keeps the analyst close to the data: the agent flags mismatches without reconciling them, a separate skill checks every figure, and each session ends with lessons fed back into the skills.
Most people copy the headline and skip those parts. Principle 26, Boredom as Resource, explains the loss. The tedious hours were where you learned how the numbers break.
Why these 5 belong together
Each story adds something to the moment you click approve. People in the next room. A promise in an old contract. Decisions inside a vendor’s plugin. Rules from a team you never met. Knowledge that only formed during the slow version.
The approval looks the same as it did in January. What it covers has grown. And a single window that decides for itself when to act makes it easier to approve without noticing what came along.
None of the 5 changes is a mistake by the vendor. Each one is useful, and several were built with care. Google keeps locked doors out of the agent’s reach. Anthropic’s report design keeps a human on every mismatch. Typeface shows which rule applied inside its own tools. The risk sits in the gap between what these features do and what an ordinary user notices at the moment of approval. That gap is where the principles in this library do their work.
Key insight: The click did not change. Its contents did. Each new connector, plugin, memory feature, and rule layer adds parties to your approval, and none of them announce themselves in the output.
Cowork task
“Before I approve the next task you complete, list everything that shaped or will be affected by it beyond my instruction in this session: people other than me, connectors or plugins that contributed, rules or guidelines from outside sources, stored memory you drew on, and any promise I have made to others that this task touches. Mark each item as visible in the output or not visible. Do not execute anything new while preparing this list.”
For individuals: Run the task above on your next important delegation. The items marked not visible are the ones your approval covers without your knowledge.
For teams: Every shared connector, plugin, and rule layer needs a named owner. Without one, each person assumes someone else checked.
For organizations: Agent products now change monthly. Treat any release that touches memory, data, or connectors as an event that needs review, the way you would treat a new vendor.
This article is part of Claude Cowork: The Tool That Shapes the Master, a living library of 30 principles and real-world examples for professionals who work with AI agents daily.
30 principles. 100+ examples. 50 workflows and automations. 99 commands. More every week. Founding members get all 30 principles and the full example library, with new examples added every week from current practice.
