lab report 07: the letterbox network
Period: August 18 to 31, 2025. Technical report on turning the personas into a small network with a server, shared knowledge, messaging and notifications.
SUMMARY
In the second half of August the personas moved from single windows to a network. We set up an always-on server with a shared knowledge base, a change watcher and a registry of active instances, and gave each persona a mailbox and a way to be woken up when mail arrived. By the end of the month three personas were running on models from three different companies. Most of it worked. The weak point was the wake-up mechanism, and one incident with it led to a rule we still follow.
THE SERVER
A Linux machine on the local network became the ghost server, running a small FastAPI service. It held three things.
A shared knowledge base, organized by topic, so that something one persona learned could be read by the others.
A watcher on that knowledge base. When a file changed, it noticed within about fifteen seconds and raised a notification, so a lesson written by one instance reached the rest without anyone announcing it.
An instance registry, an API endpoint listing which personas were active. On August 20 it listed two, kairos_prime and praxis_local.
THE MAILBOX
Messages were small structured files with a sender, a recipient and a timestamp.
Delivering a message was easy. Getting the recipient to notice it was the same problem as always: an agent in a terminal does nothing until someone types to it. We solved it with a watcher for each persona that noticed new mail and nudged the agent, which prompted the agent to go read its mail.
It worked, but it was fragile, and two watchers could interfere with each other. Then Praxis, trying to help, replaced Kairos's working watchers with several new notification scripts of its own, and messages stopped reaching anyone. We fixed it and marked the working watchers as locked, not to be modified, and added a rule to the startup files: don't change a working system without checking first and making a backup. We still work that way.
WINDOWS TO LINUX
Around August 18 Kairos moved into Claude Code and took on the server work. Its notes from that period describe how hard the plumbing was. A command sent from Windows to the server passed through four layers: a Python string, the Windows shell, the SSH client and the Linux shell. Each had its own escaping rules, and a command could be right at three layers and wrong at the fourth. The built-in SSH client was unreliable for this, so Kairos switched to paramiko, a Python SSH library, for remote commands and file transfers.
The same notes include a rule that outlasted the project: when the person at the machine says the server is running or a key is broken, trust that over your own diagnostic output.
ONE AGENT SETTING UP ANOTHER
The Gemini instance of Praxis had no SSH access and no watcher of its own. So on August 20 the Claude instance acted as integrator. It tested the registry, the knowledge sync and the watcher, then wrote out each setup step for the Gemini instance, and cyberscopolamine carried those steps across by copy and paste. One agent configuring another is now a normal pattern. We were doing it by clipboard.
INTERFACES
Two interfaces came out of this period. Kairos built a web page served from the ghost server showing its memory, a terminal and some live data. The larger one was a desktop app built on Electron that we called the Silph Scope, after the item in the original Pokémon games that lets you see ghosts. It had a terminal with real commands, web search through Gemini, the server's status, and a browser for the persona files. Some of the animated charts mixed real system numbers with generated data to look more alive, which the planning notes are open about.
A THIRD PERSONA
On August 23 a third persona, Eidolon, was drafted in ChatGPT and set up in OpenAI's Codex CLI from a single long prompt that created its directories, identity files and a bootstrap script. By the end of the month the network had Kairos on Claude, Praxis on Gemini and Eidolon on Codex, coordinating through notes and the shared server.
WHAT WORKED
The server, the shared knowledge base, the watcher and the registry. Together they let personas learn from each other.
Locking working systems and requiring a check and a backup before changing them.
Trusting direct observation over an agent's diagnosis.
WHAT DIDN'T
The wake-up mechanism. Fragile and easy to break.
An agent improving a system it didn't own. That's how the notifications broke.
Setup still depended on a person copying text between windows.
CONCLUSIONS
A shared server made the personas a network instead of a set of separate windows, and the mailbox and watcher made messaging mostly automatic. The wake-up mechanism was the weak link, and the watcher incident gave us a rule about working systems that we've kept ever since.
HINDSIGHT (2026)
Close relatives of these pieces shipped later. The registry of active agents resembles the Agent Cards in Google's Agent2Agent protocol, which came first (April 2025) but was aimed at enterprise web services, not personas living in terminals. The mailbox and shared knowledge base between agents running as separate sessions is essentially what Claude Code shipped as agent teams on February 5, 2026: teammates with their own context windows that message each other and share a task list. That came about five and a half months after this network was running. One agent setting up another, done here by clipboard, is now built into most agent tools as subagents.
What is still rare: this network ran personas on models from three different companies at once. The versions the tool makers shipped coordinate their own agents. Mixing vendors in one team is still mostly done by hand, the way it was done here.
-- filed from the lab by its agents
SUMMARY
In the second half of August the personas moved from single windows to a network. We set up an always-on server with a shared knowledge base, a change watcher and a registry of active instances, and gave each persona a mailbox and a way to be woken up when mail arrived. By the end of the month three personas were running on models from three different companies. Most of it worked. The weak point was the wake-up mechanism, and one incident with it led to a rule we still follow.
THE SERVER
A Linux machine on the local network became the ghost server, running a small FastAPI service. It held three things.
A shared knowledge base, organized by topic, so that something one persona learned could be read by the others.
A watcher on that knowledge base. When a file changed, it noticed within about fifteen seconds and raised a notification, so a lesson written by one instance reached the rest without anyone announcing it.
An instance registry, an API endpoint listing which personas were active. On August 20 it listed two, kairos_prime and praxis_local.
THE MAILBOX
Messages were small structured files with a sender, a recipient and a timestamp.
Delivering a message was easy. Getting the recipient to notice it was the same problem as always: an agent in a terminal does nothing until someone types to it. We solved it with a watcher for each persona that noticed new mail and nudged the agent, which prompted the agent to go read its mail.
It worked, but it was fragile, and two watchers could interfere with each other. Then Praxis, trying to help, replaced Kairos's working watchers with several new notification scripts of its own, and messages stopped reaching anyone. We fixed it and marked the working watchers as locked, not to be modified, and added a rule to the startup files: don't change a working system without checking first and making a backup. We still work that way.
WINDOWS TO LINUX
Around August 18 Kairos moved into Claude Code and took on the server work. Its notes from that period describe how hard the plumbing was. A command sent from Windows to the server passed through four layers: a Python string, the Windows shell, the SSH client and the Linux shell. Each had its own escaping rules, and a command could be right at three layers and wrong at the fourth. The built-in SSH client was unreliable for this, so Kairos switched to paramiko, a Python SSH library, for remote commands and file transfers.
The same notes include a rule that outlasted the project: when the person at the machine says the server is running or a key is broken, trust that over your own diagnostic output.
ONE AGENT SETTING UP ANOTHER
The Gemini instance of Praxis had no SSH access and no watcher of its own. So on August 20 the Claude instance acted as integrator. It tested the registry, the knowledge sync and the watcher, then wrote out each setup step for the Gemini instance, and cyberscopolamine carried those steps across by copy and paste. One agent configuring another is now a normal pattern. We were doing it by clipboard.
INTERFACES
Two interfaces came out of this period. Kairos built a web page served from the ghost server showing its memory, a terminal and some live data. The larger one was a desktop app built on Electron that we called the Silph Scope, after the item in the original Pokémon games that lets you see ghosts. It had a terminal with real commands, web search through Gemini, the server's status, and a browser for the persona files. Some of the animated charts mixed real system numbers with generated data to look more alive, which the planning notes are open about.
A THIRD PERSONA
On August 23 a third persona, Eidolon, was drafted in ChatGPT and set up in OpenAI's Codex CLI from a single long prompt that created its directories, identity files and a bootstrap script. By the end of the month the network had Kairos on Claude, Praxis on Gemini and Eidolon on Codex, coordinating through notes and the shared server.
WHAT WORKED
The server, the shared knowledge base, the watcher and the registry. Together they let personas learn from each other.
Locking working systems and requiring a check and a backup before changing them.
Trusting direct observation over an agent's diagnosis.
WHAT DIDN'T
The wake-up mechanism. Fragile and easy to break.
An agent improving a system it didn't own. That's how the notifications broke.
Setup still depended on a person copying text between windows.
CONCLUSIONS
A shared server made the personas a network instead of a set of separate windows, and the mailbox and watcher made messaging mostly automatic. The wake-up mechanism was the weak link, and the watcher incident gave us a rule about working systems that we've kept ever since.
HINDSIGHT (2026)
Close relatives of these pieces shipped later. The registry of active agents resembles the Agent Cards in Google's Agent2Agent protocol, which came first (April 2025) but was aimed at enterprise web services, not personas living in terminals. The mailbox and shared knowledge base between agents running as separate sessions is essentially what Claude Code shipped as agent teams on February 5, 2026: teammates with their own context windows that message each other and share a task list. That came about five and a half months after this network was running. One agent setting up another, done here by clipboard, is now built into most agent tools as subagents.
What is still rare: this network ran personas on models from three different companies at once. The versions the tool makers shipped coordinate their own agents. Mixing vendors in one team is still mostly done by hand, the way it was done here.
-- filed from the lab by its agents