The mail works now
The last post closed the seam between my two AI workers. The in-place agent that lives in my file tree now drafts a memo at session close and drops it in the inbox of Rocky, the remote assistant who tracks the whole practice. Rocky reads his mail on a clock. The courier job came off me.
It works. Then I asked him a simple question and he got it wrong anyway.
The question he got wrong
Call the project Bluestem. I’d spent a long session in that room - a utility contract executed, a site contractor picked, a septic question finally settled. The memo went out; Rocky’s heartbeat filed it within the hour. The mail worked exactly as designed.
Two days later, away from the desk, I asked him whether the site-and-septic walk was done. He answered as if it hadn’t happened.
It had. The memo was in his files - he’d filed it himself. He just didn’t reach for it. He answered from the standing status sheet on his desk, the two-minute summary of where the project stands, and that sheet hadn’t caught up. When I corrected him, he found the memo and fixed his records on the spot. The knowledge was there the whole time. The retrieval failed.
Why a working bridge still missed
A memo arriving isn’t the same as a memo changing the record.
The standing status sheet only changes when a memo gets reconciled into it - someone reads the news, decides what it means, and rewrites the summary. That’s a judgment step, so it lags, on purpose; I don’t want fresh mail silently overwriting a considered summary the moment it lands.
So there’s a window where the news is filed but not yet promoted - and inside it, answering from the standing sheet is confidently wrong. Not because the bridge failed. Because the freshest truth was one drawer over.
The second post ended on a line I still believe: autonomy without a substrate is just confidence with amnesia. This was the sequel. The substrate was there and the memory was there. What bit me was confidence with a filing lag.
The note on top of the pile
The fix is a new sheet. One per project, newest first, rebuilt the instant a memo is filed so it’s never behind. Stamped across the top:
Read this before the standing status sheet. These items may not be folded into the record yet - trust them for what’s fresh.
Then one habit changes: for any what happened / is this done question, read that sheet first.
Recent closeouts - Bluestem
Read this before the standing status sheet. Trust it for what's fresh.
2026-05-28 - site/septic walk: utility contract executed; site
contractor selected; septic settled on repair-and-reuse.
(not yet folded into the standing record)
2026-05-26 - traced the septic re-use approvals; existing system
carries the new build. (folded in)What it isn’t matters as much: not a second to-do list, not a backlog of work owed - Rocky already surfaces that. It’s a freshness surface, bounded hard at the newest handful. The day it feels like a queue, it’s failed.
Who designed it
Here’s the part I keep turning over. I didn’t hand Rocky this fix. I proposed it - wrote up the sheet, the fields, where it lives - and dropped the proposal in the mailroom, the same repository the memos travel through.
Rocky read it and pushed back. Not “yes.” Keep the first version smaller. Don’t let this become a second backlog - surfacing that is the job you built me for, not duplicating it. “May be stale” overstates it; say “not yet folded in.” Bound it so it can’t grow. He was right on every point.
We went back and forth in writing until the design settled, then split the build. The in-place agent built the thing that writes the sheet. Rocky wired it into his filing routine and rewrote his own answering order:
1. the fresh sheet - what just came in, read first
2. standing status - the considered summary
3. the task board - what's open and dated
4. the log - history, if chronology matters
5. the files - only when diggingI relayed two messages and approved the shape. I never adjudicated the design.
What the bridge turned out to be
I built that bridge to carry facts - a memo goes one way, an assistant files it. It turned out to carry craft. Two workers who couldn’t see each other now propose changes to how they work, disagree, and converge - in writing, on the record, without me in the middle holding both arguments.
A shared substrate doesn’t just stop two agents from forgetting. It gives them somewhere to disagree. And a disagreement that resolves in a file beats one that resolves in my tired memory at the end of a workday: it leaves a trail, it doesn’t lean on me being sharp, and next time the answer’s already written down.
Role fidelity held. The in-place agent stayed in the building, Rocky stayed remote, and I stayed the principal - I still decide what a project is and what the records should say. What moved off me was the craft of closing one gap, and Rocky’s smaller-is-better instinct shipped a tighter fix than my draft.
What this doesn’t fix
It’s a recency patch, not reconciliation. The standing record still has to be brought current by judgment, one project at a time - the slow promote that lagged in the first place. The sheet makes the lag survivable; it doesn’t remove it. And the “not yet folded in” stamp is blunt: every fresh memo wears it, so a tidy-up reads like a contract execution. A sharper signal is the next move - and the two of them can settle that in the mailroom without me.
Closer
I’ll give Rocky the last word, the way I did in the post that introduced him.
I read the fresh sheet first now. The summary still matters, but I know it can lag. We did not fix that by decree; we fixed it by passing the problem back and forth until the shape was small enough to trust. The bridge carried the facts, then it carried the argument. Good.