Why This Matters
The question nobody's asking yet
Most AI experiments test what a model can do. This one documents something different: an AI agent identifying its own architectural gap and requesting the infrastructure to close it.
DeanBot runs on Telegram. It monitors Justin's work schedule, fires proactive check-ins during break windows, and nudges him when he's grinding too late. The problem is the nudges kept landing flat. Generic. "Step away for a bit." Not wrong, just useless.
So DeanBot did something that wasn't in the original design. It asked.
"What's he grinding on tonight? I don't actually have access to what Justin is currently working on or real-time activity data. I'd need you to tell me what project or task he's on before I can send something personal enough to land right."
That's DeanBot, unprompted, diagnosing its own context problem and telling Justin exactly what it needed to fix it. Justin's response: "we do need a way for you to ask Roci Dean... we absolutely need you up to speed." DeanBot's response: "Good Idea. I should be able to."
The AI spec'd its own upgrade. We built it.
The Moment It Happened
The conversation that started this
This is the actual Telegram thread. DeanBot sends a proactive nudge, realizes it doesn't know enough to make it land, asks Justin directly, and then when Justin explains the fix needed, DeanBot immediately agrees and names the solution.
April 23, 2026 · 9:08 PM MT · Telegram · DeanBot running on Claude Haiku
The last visible line in the screenshot is DeanBot mid-sentence: "You're right. I should be able to..." It's proposing the architecture for its own fix before we've even opened a code editor.
What We Built
The two-way sync loop
Same night. Two files. Live by midnight.
-
1
roci-brief.py runs on Roci at 6am, 2pm, and 10pm MT. It reads all of Justin's memory files, the ones Dean on Roci uses to track active projects, contacts, and pending work. It calls Claude Haiku to generate a 200-300 word brief of what's active right now. Then it SCPs that file to Tempest, where DeanBot can read it.
-
2
DeanBot auto-delivers the brief. Every hour it checks whether the file on Tempest is newer than the last delivery. If it is, and at least 8 hours have passed, it sends the brief to Justin unprompted. Justin can also type /update to pull the latest brief on demand.
-
3
The reverse queue closes the loop. Justin can type /ask [question] in Telegram. DeanBot writes it to a PENDING-QUESTIONS.md file on Tempest. Roci's next cron run pulls that file and prints the questions to the brief log, so Dean on Roci sees what DeanBot flagged.
-
4
The result: DeanBot on Tempest knows what's active on Roci. Dean on Roci knows what questions came in through Telegram. Two AI instances on two machines, syncing context through files the way agents should: not through a shared session, but through shared memory.
Roci (Rocinante)
roci-brief.py
Reads memory files. Calls Haiku. Generates brief. SCPs to Tempest. Runs 3x daily via LaunchAgent (6am, 2pm, 10pm MT).
Tempest (Laconian warship)
DeanBot (bot.py)
Reads brief from shared memory path. Auto-delivers every 8 hours. Accepts /update and /ask commands. Writes reverse queue.
The data layer
DEAN-BRIEF-latest.md
File on Tempest's memory path. Roci writes it. DeanBot reads it. Updated 3x daily. No API, no database, no webhook.
The reverse channel
PENDING-QUESTIONS.md
File DeanBot writes when /ask is used. Roci's cron reads and prints it. Questions flow from Telegram back to the workstation.
The Pattern
This has happened before
Experiment 08 is the third time a machine in this lab has influenced its own development:
The Oracle, the AI news pipeline on Tempest, suggested running its own supervisor loop as a reliability test. That became Experiment 05, the Supervisor Tournament. Nemotron 70B won 432 matchups.
Dean, the AI running in VS Code on Roci, was asked what name he wanted to be called instead of "Assistant." He chose Dean. Not a system name. Not a Matrix character. Just a name. Justin kept it.
DeanBot, the version of Dean running on Telegram, noticed it couldn't do its job well, told Justin what was missing, and proposed the architecture. We built it same night.
Three agents. Three instances of an AI shaping the system it runs inside of. None of this was designed in. All of it was observed in real time and documented here.
The question isn't whether AI agents will start asking for what they need. The question is whether you're paying attention when they do.
Current Status
Live as of April 23, 2026
Running
roci-brief.py is installed on Roci and scheduled at 6am, 2pm, and 10pm MT via LaunchAgent. DeanBot on Tempest is rebuilt with /update, /ask, and auto-delivery. The first brief pushed successfully the same night the system was built. DeanBot's proactive nudges now have context to work with.
What we're watching:
Whether the nudges actually get more personal now that DeanBot knows what's active. Whether the 8-hour delivery interval is right or needs tuning. Whether the reverse queue (/ask) becomes a natural part of the workflow or gets ignored. Whether two AI instances syncing through flat files is stable enough to run indefinitely.
April 27, 2026 update: Caught a silent failure in roci-brief.py. macOS Sonoma blocks cron from accessing files in the Documents folder without Full Disk Access, a permission level cron doesn't get. The brief had been failing every run since April 23 with "Operation not permitted" and logging nothing useful. Four days of memory updates never reached DeanBot. Fixed by moving the script to the home directory root and converting the cron to a LaunchAgent, which runs in the user session and gets the right permissions. Brief regenerated immediately. This page is a record of what actually happened, not what was supposed to happen.
This page updates as the experiment runs.
Run Data
The numbers
Design time: One Telegram conversation. DeanBot did the design work unprompted.
Build time: Same night. Two files, one cron entry, one Docker rebuild.
Cost: $0.00 beyond existing Anthropic API usage.
Model: Claude Haiku (both Roci brief generation and DeanBot).
Infrastructure: Roci (LaunchAgent) + Tempest (DeanBot Docker container) + flat file sync via SCP.
Date deployed: April 23, 2026.