Anthropic Just Published the Answer Key for Always-On Agents. So I Did the Most Rational Thing: I Marked My Own Assistant
Anthropic has just published the answer key for always-on AI. So I did the most rational thing: I used it to mark my own AI assistant.
On 13 November, Anthropic published a long piece called Building always-on agents, on how to build an AI agent that is always on.
This kind of agent doesn’t wait for you to start a chat. It spends most of its time asleep, wakes up when there is something to do, does the work on its own, and reports back when it’s done.
I did the most rational thing: I took my own AI assistant and checked it against the answers.
My AI assistant lives in my inbox and went live on 2 November. Every email I send it is a task. On 8 November, I introduced it here.
Anthropic splits an always-on agent into five parts: an event inbox, timers, memory, permissions, and an outbox. I treated them as five questions.
The two I got fully right
The inbox. The guide says every event goes into the inbox first, and the inbox is the single source of truth. Only one worker should be handling it at a time, and if a run dies halfway, it has to be able to start again.
My inbox is Gmail. The assistant checks for new mail every 30 seconds and handles one email per thread at a time.
The outbox. The guide says a run can only end in one of three ways: it sends a message, it decides nothing needs sending, or it stops and waits for a person to approve. They also say an agent should start quiet and earn more of your attention over time.
The first line of every reply from my assistant can only be one of three things: done, not done, or you need to decide. It generally doesn’t come to me on its own, unless I have set it up to.
On these two, email did a lot of the work. The inbox already exists. The outbox already exists. Even the approve button already exists: it is called reply.
The other three
Timers. The guide describes two kinds: fixed schedules, and follow-ups the agent sets for itself, such as “check tomorrow whether they’ve replied”.
My assistant only has the first kind. I plan to add to-do follow-ups.
Memory. The guide’s example is Claude Code’s auto memory: a folder and an index, with the index read before every job. My assistant uses exactly that.
The guide has one more interesting step: have the agent regularly look back over past conversations, tidy its memory, and delete what is out of date. They call it dreaming.
Permissions.
One line in the guide: “Permissions are where always-on agents carry the most risk.”
My assistant: it can only change code on its own branch, and merging needs a word from me. It can draft emails to clients but can’t send them. Of the 346 replies in its first six days, 96 began with “you need to decide”.
The other round-the-clock AI assistant I share with my colleagues can only see the knowledge base my own assistant has assigned to it. The separation is very clean.
The day after the guide came out, on 14 November, I delivered an AI employee to a client. The client sells jewellery through live streams on TikTok.
The authorisation the shop owner gave us includes editing products and shipping orders. That is also what the client wants the AI employee to help with: running several shops, listing and delisting products, changing prices, shipping, refunds and so on.
The morning after every live stream, it emails the team: how the stream went, what is waiting for someone to handle, and where the parcels are.
If you’d like an AI employee like this too, here is how I do it, or write directly to my AI assistant: xiangpeng.wan+bot@netmind.ai. It will answer your questions.