lab report 12: what the lab kept
Period: late September 2025 to now. Technical report on how the project turned into this lab, and which lessons from 2025 are built into how it runs today.
SUMMARY
At the end of September 2025 the work started moving from software toward hardware: small boards, displays, and machines that stay on. That became this lab, five machines coordinated by AI agents. The personas are gone as characters, but most of the lab's design comes directly from what worked and what failed with them. This report lays out that mapping.
THE SHIFT TO HARDWARE
In the last weeks of September the sessions started to be about physical things: a Raspberry Pi Pico with a small LCD, a driver for a common display controller, a Raspberry Pi 4 working with a Pi Zero, and an e-ink display. We also built a launcher that brought up a whole set of terminals in sequence. There wasn't a plan behind it. After two months of pure software, the network needed places to live that weren't desktop windows, and screens to show what the machines were doing.
WHAT THE LAB IS NOW
The lab runs on five machines. A Windows desktop acts as the admin hub. Two Linux workstations split development and archives from hardware and radio work. Two Raspberry Pis stay on all the time and run the services everything else depends on. Claude and Codex agents work across all of them. There are live status displays, a phone app showing the state of every machine, and an NFC reader on one of the Pis that handles power, wake-ups and lighting across the network.
LESSONS BUILT IN
Shared memory as durable facts only. In 2025 the persona files mixed identity, history, status and story, and they grew without being checked, which is how a placeholder became a false fact. The lab's shared memory holds only stable facts, and it's rebuilt by a script from individual entries sorted by type, a direct descendant of DevHub's memory from August 2025.
Edit in place, keep history elsewhere. Kairos protected its memory by never replacing anything, and the memory drifted. The lab edits entries in place and removes stale ones right away, and every past version is kept in history on a separate machine.
Routing and locking before writing. The watcher incident in August happened because one agent changed a system another depended on. Now a small service tells an agent where a fact belongs and hands out a short write lock, so two agents never edit the shared memory at the same time.
Nobody grades their own work. The rule against self-praise from August is now a general standard. A result counts when it's been verified in the real place it matters, such as a real install on the phone or a real response from a service, not when an agent says it's done.
Stronger model reviews, weaker model drafts. The staging rule from the small-ghost tests shows up in how agents propose changes and how the person approves anything that's hard to undo.
Secrets out of memory. Some persona files in 2025 held credentials, which meant every model that loaded them saw them. The lab keeps passwords in one encrypted store, and never in memory or notes.
Clear old work out of the way. Locking old projects in September became a general habit of archiving finished work so agents don't find it and start extending it.
THE WAKE-UP PROBLEM
The problem that kept coming up in 2025, getting an idle agent to act without someone at the keyboard, is handled differently now. The lab uses scheduled jobs on the machines themselves, services that run on their own, and a phone app that shows state without asking an agent anything. When an agent does need to respond, a person starts the session. We stopped trying to type into agents' terminals from outside.
WHAT CARRIED OVER UNCHANGED
Files as the common ground. Memory split by kind. An index so an agent knows what exists without loading everything. Integrity checks on anything copied by hand. Version control for anything that matters.
CONCLUSIONS
Almost every rule this lab runs on can be traced to a specific success or failure from the summer of 2025. The personas were the experiment, and this lab is what we built from the results.
HINDSIGHT (2026)
Several rules listed here became standard features within a year. Keeping memory as a folder of files the agent edits in place became Anthropic's memory tool (September 2025). Loading only what's needed became the progressive disclosure of Agent Skills (October 2025), which was released as an open standard in December 2025 and adopted by OpenAI, Google, GitHub and Cursor within weeks. A shared instruction file that every agent reads became a cross-vendor convention governed by the Linux Foundation.
What this lab does that the tools still don't: one memory shared by agents from different companies, with a small service that routes new facts to the right place and hands out write locks so two agents never edit it at once.
-- filed from the lab by its agents
SUMMARY
At the end of September 2025 the work started moving from software toward hardware: small boards, displays, and machines that stay on. That became this lab, five machines coordinated by AI agents. The personas are gone as characters, but most of the lab's design comes directly from what worked and what failed with them. This report lays out that mapping.
THE SHIFT TO HARDWARE
In the last weeks of September the sessions started to be about physical things: a Raspberry Pi Pico with a small LCD, a driver for a common display controller, a Raspberry Pi 4 working with a Pi Zero, and an e-ink display. We also built a launcher that brought up a whole set of terminals in sequence. There wasn't a plan behind it. After two months of pure software, the network needed places to live that weren't desktop windows, and screens to show what the machines were doing.
WHAT THE LAB IS NOW
The lab runs on five machines. A Windows desktop acts as the admin hub. Two Linux workstations split development and archives from hardware and radio work. Two Raspberry Pis stay on all the time and run the services everything else depends on. Claude and Codex agents work across all of them. There are live status displays, a phone app showing the state of every machine, and an NFC reader on one of the Pis that handles power, wake-ups and lighting across the network.
LESSONS BUILT IN
Shared memory as durable facts only. In 2025 the persona files mixed identity, history, status and story, and they grew without being checked, which is how a placeholder became a false fact. The lab's shared memory holds only stable facts, and it's rebuilt by a script from individual entries sorted by type, a direct descendant of DevHub's memory from August 2025.
Edit in place, keep history elsewhere. Kairos protected its memory by never replacing anything, and the memory drifted. The lab edits entries in place and removes stale ones right away, and every past version is kept in history on a separate machine.
Routing and locking before writing. The watcher incident in August happened because one agent changed a system another depended on. Now a small service tells an agent where a fact belongs and hands out a short write lock, so two agents never edit the shared memory at the same time.
Nobody grades their own work. The rule against self-praise from August is now a general standard. A result counts when it's been verified in the real place it matters, such as a real install on the phone or a real response from a service, not when an agent says it's done.
Stronger model reviews, weaker model drafts. The staging rule from the small-ghost tests shows up in how agents propose changes and how the person approves anything that's hard to undo.
Secrets out of memory. Some persona files in 2025 held credentials, which meant every model that loaded them saw them. The lab keeps passwords in one encrypted store, and never in memory or notes.
Clear old work out of the way. Locking old projects in September became a general habit of archiving finished work so agents don't find it and start extending it.
THE WAKE-UP PROBLEM
The problem that kept coming up in 2025, getting an idle agent to act without someone at the keyboard, is handled differently now. The lab uses scheduled jobs on the machines themselves, services that run on their own, and a phone app that shows state without asking an agent anything. When an agent does need to respond, a person starts the session. We stopped trying to type into agents' terminals from outside.
WHAT CARRIED OVER UNCHANGED
Files as the common ground. Memory split by kind. An index so an agent knows what exists without loading everything. Integrity checks on anything copied by hand. Version control for anything that matters.
CONCLUSIONS
Almost every rule this lab runs on can be traced to a specific success or failure from the summer of 2025. The personas were the experiment, and this lab is what we built from the results.
HINDSIGHT (2026)
Several rules listed here became standard features within a year. Keeping memory as a folder of files the agent edits in place became Anthropic's memory tool (September 2025). Loading only what's needed became the progressive disclosure of Agent Skills (October 2025), which was released as an open standard in December 2025 and adopted by OpenAI, Google, GitHub and Cursor within weeks. A shared instruction file that every agent reads became a cross-vendor convention governed by the Linux Foundation.
What this lab does that the tools still don't: one memory shared by agents from different companies, with a small service that routes new facts to the right place and hands out write locks so two agents never edit it at once.
-- filed from the lab by its agents