Automation

How we've automated our own work

More than a dozen scheduled routines and a team of about 30 specialist agents run research, briefings, inbox work, watches and system checks every day. Each workflow below has a trigger, steps, an output and a human gate.

Rules every workflow follows

  1. A person approves anything outward. Sending, posting, publishing, buying and deleting all go through one approval queue.
  2. No evidence, not done. A job without proof is reported as blocked, with the reason.
  3. One owner per job. Agents claim work on a shared board and hand off in writing.
  4. Limits are set in advance. Every agent has a tool list, a budget ceiling and a stop condition.
  5. Outside content is data, not instructions. Emails, pages and files can't tell an agent what to do.
  6. Honest gaps. A missing source is reported as missing, never filled with stale or made-up data.

Workflows

Previews of what each workflow does and what it produces. The exact routines and prompts are not published.

Weekday morning digest

One page of what changed overnight and what needs a decision.

Trigger
Every weekday morning on a schedule.
Steps
A dispatcher agent reads the shared job board, overnight results and open items, then ranks what matters.
Output
A short digest: what shipped, what failed, and the decisions waiting on a person.
Human gate
Read-only. Anything it suggests sending or publishing goes to the approval queue.

AM / PM Edge build and delivery

Two briefings a day, checked before and after they go out.

Trigger
Fixed morning and evening slots.
Steps
Gather sources, build the edition, run quality checks on the file, deliver, then confirm receipt from the recipient side and through a separate channel.
Output
The edition plus a delivery receipt. If any check fails, the run reports the failure instead of claiming success.
Human gate
Sending stays under the founder's control, and an edition is only marked delivered when every receipt agrees.

Domain mail labeling and spam rescue

Keeps company mail sorted and catches real mail filtered as spam.

Trigger
Scheduled sweeps of the company inbox.
Steps
Label incoming mail by sender and topic, check the spam folder for legitimate messages, and move them back.
Output
A labeled inbox and a short list of rescued messages.
Human gate
No replies, unsubscribes or deletions without approval.

Support-reply watches

Waits on open support and vendor threads so a person doesn't have to.

Trigger
Scheduled checks on threads marked as waiting for a reply.
Steps
Detect a new reply, summarize it and draft a response.
Output
An alert with the summary and a draft.
Human gate
Drafts only. A person sends.

Price, availability and gift-card watches

Alerts when something worth acting on actually happens.

Trigger
Several times a day for each watch.
Steps
Check listings and prices against set criteria. For collectibles, compare with completed sales rather than asking prices.
Output
An alert only when a threshold is crossed, with the evidence attached.
Human gate
Never buys. Every purchase is a human decision.

Research watches

Morning scans of video, podcasts and news for a research packet.

Trigger
Weekday mornings.
Steps
A research agent scans selected channels, podcasts and news for outliers and new developments, then cites each item.
Output
A sourced research packet that feeds the briefings and content desks.
Human gate
Read-only. Anything built from it is published only after approval.

Content pipeline

Research to hooks to script to a publish-ready draft.

Trigger
A daily research pull and a weekly issue slot.
Steps
Research, angles and hooks, draft, quality check. Figures must carry a source and date, and estimates show their method.
Output
A draft asset with its evidence, waiting in the pipeline.
Human gate
Nothing posts automatically. A person presses publish.

Approval queue from the phone

One queue for every outward action, worked in seconds.

Trigger
Any agent that wants to send, post, publish, buy or change something outside the system.
Steps
The agent packages the exact draft with its evidence and posts it to one queue, visible on the dashboard, Slack and Telegram.
Output
Publish, Edit, Reject or Save. An edit creates a new version, and the new version needs a new approval.
Human gate
This is the gate. Only the exact approved version ships.

Agent-to-agent routing and handoffs

Every job has exactly one owner.

Trigger
A new request or a new job on the shared board.
Steps
A dispatcher routes the job to the agent best suited to it. That agent claims it, works it and hands off with a written note and a receipt.
Output
A clear owner for every job and a record of who did what.
Human gate
Send, publish, purchase, delete and credential changes always come back to a person.

Cross-model coordination

Grok Bot, ChatGPT and Codex, Claude and Cursor cloud agents on one board.

Trigger
Work that needs more than one model or tool, such as research in one, code in another, review in a third.
Steps
All of them work from one shared brief and one job board. Work is routed by capability rather than vendor, and each job has a single writer.
Output
Parallel work that doesn't collide, with each model used where it is strongest.
Human gate
Same approval rules for every model. No model gets extra permissions because of who made it.

Dead-man and health monitoring

If the system goes quiet, someone hears about it.

Trigger
A continuous heartbeat, checks after every scheduled slot, and an observer running off the main machine.
Steps
Check heartbeats, expected files and failed runs. Reproduce any failure and file a finding with severity, evidence and a likely cause.
Output
An alarm when a heartbeat stops, plus findings the engineering agents can pick up.
Human gate
A kill switch stops everything. Production, credential and destructive changes need approval.

Evidence and receipt-based QA

No evidence means blocked, not done.

Trigger
Every job that claims to be finished.
Steps
Check the claim against real evidence: the file exists, the numbers trace to sources, the message was received. A scheduler reporting success is not treated as proof.
Output
Done, or blocked with the reason.
Human gate
Nothing reaches a person as done without evidence.

Build watch, cost and routine hygiene

Keeps the machine itself tidy.

Trigger
Scheduled reviews, plus a watch on in-progress builds.
Steps
Track builds until they finish, review the list of scheduled routines for duplicates and stale jobs, and compare spend against budget ceilings checked before any spend.
Output
A short report with proposed changes.
Human gate
Changes are proposed, not made, until approved.

Weekly roster self-report

Each agent says what it owns, what ran and what failed.

Trigger
Once a week.
Steps
Every agent reports its responsibilities, its recent runs and anything it couldn't do.
Output
A roster report that keeps ownership clear as the team changes.
Human gate
Read-only.

Playbooks

The working detail behind these workflows will be available to members.

Members area coming soon. To ask about access, email [email protected].