Notes · 19 July 2026
What changed in Claude this month, and what it means for a working business.
Part of what a client pays us for is that we keep up, so this is the first of a recurring note: the month's Claude changes, filtered to the ones that matter if you run a business on it. Everything below is from Anthropic's own release notes, mid-June to mid-July.
Cowork now runs on web and mobile
Cowork, the working environment our setups are built around, is no longer tied to the desktop app. Sessions can now run remotely: they are saved to your Claude account, carry across devices, and keep working after you close the laptop. Scheduled tasks run with no device online at all. It is rolling out gradually, starting with the Max plan.
What it means: the piece of this stack that felt most like software installed on one machine is becoming infrastructure. A morning routine that summarises the inbox, or a weekly report that assembles itself, no longer depends on your computer being awake. If you held off on scheduled work for exactly that reason, this is the change that removes the excuse.
Memory was rebuilt
Claude's built-in memory no longer works as a daily summary. It now keeps individual, categorised entries that it reads and updates as you talk. This is the feature closest to what we sell, so it deserves an honest reading: recall of the gist is getting better, and each round of improvement makes casual use smoother.
The distinction that matters has not moved, though. Memory is still something Claude writes for itself, automatically, by its own judgement of what mattered. Your returns window, your supplier terms, the exception you agreed in March and the way you sign off an email are structure, and structure belongs in files you own: written once, versioned, corrected when the business changes, read at the start of every session. The better memory gets, the more it helps; it does not replace the written layer any more than a colleague's good recall replaces the policy document.
Microsoft 365 learned to write
The Microsoft 365 connector was read-only. It now has write tools: with them enabled, Claude can draft and send email, manage calendar events, and create and update files in OneDrive and SharePoint. Two important limits: a Microsoft administrator has to consent to the new permissions and switch them on for the organisation, and Teams stays read-only.
What it means: businesses on the Microsoft side just caught up a long way, and the question shifts from what Claude can reach to what it is allowed to do unattended. Our position has not changed with the feature: reach is welcome, and every action that matters still belongs behind an approval gate. That a system can send is not a reason to let it.
Two new models, one bumpy launch
Claude Sonnet 5 arrived at the end of June, the most capable model in its class for tool use and agent work. Above it, Anthropic launched Claude Fable 5 in early June, a new top tier, then suspended access for three weeks and restored it at the start of July.
What it means: the models will keep churning, and a business should feel almost none of it. A system built on written context carries across model swaps unchanged; the files that describe your business do not care which model reads them. If your Claude setup would break when the model changes, the setup is the problem. The month's small quality-of-life change points the same way: drafts can now be edited in place, so you mark the paragraph you want changed instead of re-describing it in chat.
The pattern
Every month the platform absorbs more of the mechanical work: the install, the reach into your tools, the recall of the gist. What it does not absorb is your business's own structure, because only you can write that down. That layer is the durable one, and it is the one you own.
If you want to build it yourself, the free Context Kit is the folder structure we use, sent by email. If you would rather build it with someone, the 60-minute Setup Session installs Claude and writes your first context file inside the hour.
Daniel · The Clarion