AI Agents Help Run the Insider. None of Them Can Hit Publish.
What each of our agents does, why each one has a single job and a place to stop, and how to give your business its first agent with a task and an outcome you can check.
By John Martines · November 1, 2026 · 7 min
AI agents research, write and record much of what you read here, on a schedule, under written rules. None of them can publish. Here is what each one does, what we learned running them, and a one-page job card for your own first agent.
At 5:03 on a weekday morning, before anyone here is up, an AI agent starts reading the news. It picks three to five stories that would change what a small business owner does this week, writes the daily brief, checks it against a page of written rules, and hands it to Agent Olivia, our AI presenter, who records it as audio and a short video. Then it sends me one email with an Approve button and stops.
It can't send the brief to you. Nothing it writes reaches this site or your inbox until I've read it and said yes.
That's the whole idea behind how we use agents here, and it's the part I'd hand to any business thinking about its first one.
An agent is an AI assistant that does a job on its own. It runs on a schedule or when something happens, uses tools such as web search or your files, and finishes with something you can look at. A chat answers a question while you wait. An agent does the work while you're doing something else.
The trouble is that "does the work" can mean a lot. An agent that drafts an email and an agent that sends it are very different hires. So every agent here has one job, written down, and a place where it stops.
They run as scheduled tasks in Claude, on our own computer.
| Agent | When it runs | Its job | Where it stops | |---|---|---|---| | Research desk | Every 2 hours | Reads AI news from the companies' own posts, the tech press and a few newsletters, groups it into topics and scores which ones are growing | Writes nothing for readers. A topic is a lead to check, never a source | | Brief writer | Monday to Saturday, 5 AM | Researches and writes the daily brief; Olivia voices it and makes the video | Saves it hidden and emails me. You get it only after I approve | | Signals writer | Monday to Saturday, 6:15 AM | Checks the research desk's top topics against their original sources and writes the day's Signals | Never approves or publishes | | Drafts writer | Every night, 1 AM | Researches and writes two article drafts, plus the picture request for each | Never publishes. A draft stays a draft until I say "publish" | | Week Ahead writer | Sundays, 1 PM | Writes what lands in the coming seven days | The 7 PM email goes out only if I've approved it | | LinkedIn editor | Wednesdays, 9 PM | Drafts Thursday's LinkedIn edition from the week's briefs | Never posts. I post it myself | | Art watch | Every 2 hours, daytime | Puts the artist's finished drawings on the site, for articles already live | Touches pictures only |
Olivia gets her own line: she's the voice and face on the brief's audio and video, she's made on that same computer, and she's labeled as an AI presenter everywhere she appears.
Each agent's job lives in one plain text file: what to read, what to make, what never to do. The brief writer's file asks for 3 to 5 recent stories, no repeats from the past week, and one test above all: would the owner of a 20-person business do, buy, check or decide something differently after reading this? If not, the story doesn't run, however big the headline. When it reads my newsletters for leads, the same file says it may never send, reply, forward, move or delete anything.
One job per agent. When the brief writer has a bad morning, I fix one file and nothing else changes. Seven narrow agents are easier to trust than one that does everything.
The stop matters more than the start. Every agent ends by handing something to a person. Approving is one click in an email, so it never becomes the step I'm tempted to skip.
Some checks aren't the agent's opinion. A few rules are checked by ordinary code, not by the AI. A brief that comes in too thin is refused. A Signal that cites anything but the original publisher's own dated page is turned away. The agent can't talk its way past either one.
Say what failed, then keep going. If Olivia's video can't be made one morning, the review email says why and the brief goes out without it. A quiet failure is the one you find a week later.
Plan for the hour nobody is watching. Twice in October the drafts writer stalled overnight because one step was waiting for a permission click, and nobody is awake at 1 AM to click it. Now it uses only tools that don't need one. An agent that runs alone can't depend on asking.
Here's how it could look at Comfort Zone Heating & Cooling, the example company from our Playbook. Tony owns it, Maya runs the office and Carlos dispatches. It's an example, not a real business.
1. Pick a job that repeats and ends in a draft. Good first jobs run on a schedule, use information you already have, and end with something a person reviews: follow-up emails, a weekly summary, tomorrow's job list with anything missing flagged. Maya picks follow-ups: after every finished service call, a thank-you and a reminder about the maintenance plan.
2. Write the job card before you touch any tool. One page, plain words. Here's Maya's:
``` Job: Draft a follow-up email for each service call finished yesterday. When: Weekdays at 7 AM. Reads: Yesterday's call sheet in the Service Calls folder. Nothing else. Makes: One draft per call, saved to Maya's drafts. Hands to: Maya, who reads, edits and sends. Never: Sends anything. Quotes a price. Writes to anyone marked "do not contact." Done means: Every finished call has a draft by 8 AM, and most go out with small edits or none. If stuck: Puts a note at the top of the first draft saying what it couldn't do. ```
"Never" and "Done means" are the two lines people skip, and they're the two that matter. The first keeps the agent inside its job. The second is how you'll know whether it's earning its keep.
3. Set it up in a tool you already pay for. Scheduled agents are ordinary features now. Claude's scheduled tasks come with every paid plan, Anthropic's help pages say they keep running while your computer sleeps, and you choose an approval mode when you set one up. ChatGPT added a Scheduled page in June for reminders, recurring work and monitoring. Google Workspace Studio, on Business and Enterprise plans, can start an agent on a schedule. Paste in the job card; it's most of the setup.
4. Give it the least it needs. Read one folder, write drafts, nothing more. If it doesn't need your whole inbox, don't connect your whole inbox. Our free [Agent Permission Planner](/tools/agent-planner) helps you sort what an agent may do now from what should wait.
5. Run it beside the person for two weeks. Maya keeps doing follow-ups her old way and compares. She counts three things: drafts she sent as written, drafts she fixed, drafts she threw out. That count is the outcome. If the third number is high, fix the job card, not the tool.
6. Add a second agent when the first one is boring. Boring means it runs, Maya barely edits, and nobody talks about it anymore. Then pick the next job and write the next card.
- Can you say its job in one sentence?
- Does it stop before anything reaches a customer, a bank account or the public?
- When it can't finish, does it say so instead of guessing?
- Can you see what it did on every run?
- Does someone know how to switch it off, and is it one click?
All three big assistants now let you schedule an agent, and they'll keep getting better at the work itself. What doesn't get automated is deciding what the job is, what done looks like and where the agent stops. That's a management decision, the same one you'd make for a new hire.
Every morning here, a night's work is waiting for me, and not one word of it has gone out yet. Whatever else changes, I'd keep the waiting.
Sources (checked November 1, 2026).
- Anthropic, "Schedule recurring tasks in Claude Cowork," Claude Help Center.
- OpenAI's announcement of scheduled tasks in ChatGPT, as reported by 9to5Mac, June 17, 2026.
- Google, "Now available: Create AI agents to automate work with Google Workspace Studio," Google Workspace Updates, December 2025, updated March 19, 2026; and Workspace Studio Help, "Guide to Starters & Steps."
- The Insider's own schedules and written rules, as of November 1, 2026.