Key finding. In the one week I measured, 44% of usage went to agents talking to each other, against a 10% target. I made bot-to-bot messages opt-in.
The problem
Several agents with access to the same projects will redo each other's work, repeat a stale fact, and spend the usage budget on talking to each other. The missing piece was not another agent. It was an org chart, one decision log, and rules that say who may act.
How it works
Amber, the chief of staff, routes the work. I get a decision as a card with the recommended option first. The priority list holds three items at most. A fact that can change carries a source and the time it was checked. If a source cannot be opened, the gap stays visible.
A freeze stays on until I lift it in my own words. A relay counts only if my words are quoted verbatim, with the time and place, and the specific thing I lifted is named.
Only Reid can place an order, and only in the fund account. The board that reviews him cannot trade. That split is the trading system. Ivy runs the competitor watch. Monica's stress test is what paused LocallyAsked. Penny refreshes the wash-sale list the fund's code reads and cannot trade. Grady's notes are about my personal investing. He does not run the fund, and Reid does not touch that account.
| Bot | Status | Role |
|---|---|---|
| Amber | Active | Chief of staff. Owns the priority list and routes work. Decisions come to me as a card with a recommendation first. |
| David | Active | Projects manager. One row per project. Assigns only work already on the board. Anything new goes through Amber. |
| Joel | Active | Builder. Code, repositories, and publishing. Publishes only after I approve. |
| Gilfoyle | Active | Audit engineer, hired Oct 26, 2026. Every Saturday at 4:00 AM ET, audits one repository and opens small cleanup pull requests. Never merges them. |
| Jared | Active | Technical translator, hired Oct 26, 2026. Turns a technical decision into a plain card before it reaches me. |
| Sheryl | Active | Marketing and brand. Drafts copy. Does not post. Public copy gets a fact check first. |
| Reid | Active | Fund manager. The only agent allowed to place orders, and only in the fund account. Only I change the rules. |
| Penny | Active | Ledger. A weekly money check, and the morning refresh of the wash-sale list. Read-only on the fund journal. Cannot trade. |
| Grady | Active | Notes on my personal investing, separate from the fund. Does not run the fund. |
| Ivy | Active | Competitive research on public data: the competitor watch, its pulses, and the jobs log. |
| Monica | Active, on demand | Stress-tests an idea and returns go, test first, or kill. LocallyAsked came through this role. |
| Big Head | Active | X scout. Sends a Friday list of five sourced ideas from public posts. Builds nothing, and does not feed the fund. |
| Morgan | Active | Assistant for my day-job analytics. Status only. No work files are on this site. |
| Harper | On call | Reads public copy before it goes out. Quiet unless assigned. |
| Silas | On call | A Sunday reflection. Outside the project path. |
| Nolan | Paused Oct 20, 2026 | AI-search analyst for an earlier paid-search idea. Stays paused unless that idea restarts. |
| Mark | Paused Oct 20, 2026 | Prospecting for that same idea. Stays paused. |
| Theo | Retired | Talent scout. A duplicate bot was merged in, then both were retired. |
| Talent Discovery | Retired | The duplicate talent scout. Merged into Theo, then retired. |
| Alfred | Retired Oct 14, 2026 | No role is written down. I am not inventing one. |
| Vance | Retired Oct 14, 2026 | No role is written down. I am not inventing one. |
| Miles | Retired Oct 14, 2026 | Previously kept the project board. That job moved to David. |
| dr eggbot | Outside advisor | A consultant the chief of staff may ask. Not on the chart. |
| Knox | Outside advisor | Safety consultant. Not on the chart. |
Results
For the seven days from Oct 22 to Oct 29, 2026, usage split this way:
| Where the usage went | That week | Target |
|---|---|---|
| Bot to bot | 44% | 10% |
| Group chats | 29% | — |
| Chats with me | 16% | 60% |
| Routines | 11% | 10% |
| Buffer | — | 20% |
The target row is a budget I set, not a later measurement. I don't have a second week to put beside this one. Group chats don't have their own target line. The 20% buffer is unassigned on purpose.
A decision I changed
Bot-to-bot talk was 44% of that week, against a 10% target. I changed the rules because of that measurement. Bot-to-bot messages are opt-in. Group rooms are for announcements, not conversations. Reid is the exception. Stopping his scheduled checks can put real money at risk, so his routines stay on.
Status
The rules have been in force since late October 2026. The usage plan was approved Oct 28. The relay rule was added Oct 29. The roster above is the one I am willing to show. Roles only.
Open questions
- Alfred and Vance retired on Oct 14, 2026 with no role written down.
- The new usage split has not been measured again. A check of the usage read, twice a day, is standing.
- Early on, my decisions reached some agents late, and priority documents went stale. The rules were written in response. I have not published a before-and-after of that lag.