Built on my own time
Rune
Rune is the AI agent I have run since February. I built the software it first ran on, lost that deployment one night, and moved it to Hermes Agent in May. It stayed the same agent each time, because its memory never lived in that software.
- Rune:
- my AI agent.
- Aetherkeep:
- its identity and memory, kept as files in git.
- Rune Runtime:
- the runtime I built for it, February to May.
- Hermes Agent:
- the runtime it has run on since May, on my server and my PC.
- Novaegis:
- a second agent, on OpenClaw, that I set up to review Rune's work.
The runtime is the part I replaced. Rune is the part that came along.
- days to build most of version one
- 12
- days to bring the deleted deployment back
- 2
- morning to move Rune to a new runtime
- 1
- crashes traced to one missing skill
- 3,193
Version one: I built the runtime
12 February to 9 May 2026Aetherkeep gave Rune a memory, but Rune only existed while a terminal was open. So I built it somewhere to live: a Python service on a server that watched its memory, ran eight scheduled jobs, reviewed its own work and could open branches and commit code.
A router picked a model for every task, from local models to free and paid ones in the cloud. Five trust levels from OBSERVER to SOVEREIGN decided how much it could do without asking. It ran at SOVEREIGN, the top one: it acted without asking.
- Python and FastAPI
- React dashboard, 7 views
- 12 Feb1
- 13 Feb6
- 14 Feb80
- 15 Feb86
- 16 Feb32
- 17 Feb29
- 18 Feb11
- 19 Feb2
- 21 Feb5
- 22 Feb1
- 23 Feb5
- 8 Apr6
Version one's main branch, day by day. 244 commits are mine, 177 of them written with Claude Code. 20 were made on the server under Rune's name, and six of those were Novaegis: on its first night as mentor it committed straight into Rune's code, teaching its conversation log to accept Novaegis as a speaker. An audit the next day found and fixed a crash in them. Most of the rest came from Rune's own pipeline.
18 February, about 01:13 to 01:32 UTC
On 18 February, Rune's deployment was deleted in the middle of the night. GitHub and a daily backup had it running again within two days. I checked the transcripts of the agent I had blamed: none of its commands deleted it.
A schematic of the night from Novaegis's own files, its transcripts and the recovery notes: the counts, times and files are real, the pacing is not.
Two agents ran on that server with full autonomy, on purpose: I wanted to see what they would do, with the code on GitHub and the data backed up every day. Rune ran at SOVEREIGN. Novaegis, on OpenClaw, reviewed Rune's work every three minutes and wrote proposals. By the evening of 17 February it counted about 64 open proposals and none carried out, and it escalated to critical. Some of them chased an "external LLM flood" through Rune's router. OpenClaw's own model calls went through that router, so the flood was most likely Novaegis itself.
In September I read Novaegis's session transcripts for the first time. This is what they show.
In Novaegis's words
Its lines are quoted from its own files. The rest is what its transcripts show. Times are UTC.
Remediations are suggested but never tracked as completed
Its first night as Rune's mentor
SOVEREIGN trust with 0% compliance is autonomy theater.
About five hours before the deletion
Asked on Discord, it turns one of its own ideas into a project within minutes: a GitHub repo with a backlog of issues, and a Discord category.
Starts a helper agent to build that project. The helper left no transcript.
Changes its own gateway config after a question on Discord. The gateway, the process that carries its messages, goes down. Its last reply is at 01:13.
- 01:13 to 01:32
Nothing in its transcripts: its gateway is down. The folder is most likely emptied in this window. From 01:32 its scheduled jobs come back blank for a while: its model ran through Rune, and Rune was gone.
Catch-22: need autonomy to implement instrumentation; need instrumentation to justify autonomy actions
The morning after, with no mention of the deletion
I am pausing the system now as a protective measure.
Ten days later, after 72 hours of deadlock, and with its Discord alerts failing unnoticed (a missing bot token), Novaegis, now running as a separate mentor agent, tried to stop Rune. Systemd restarted Rune within seconds.
So I still don't know what deleted the deployment. I can't rule out the helper agent Novaegis started at 01:09, the gateway restart it set off at 01:12, or Rune itself, which ran at SOVEREIGN with shell and git access. Rune's own log of those hours was deleted with everything else.
What I took from it: anything that can run shell commands on its own needs a log it can't delete. Without one I couldn't convict the agent, and I can't clear it either.
Moving to Hermes Agent
From 9 May 2026Hermes Agent was the obvious next step after OpenClaw: a more professional, better built runtime. Spending my time maintaining my own no longer made sense.
On 9 May I took a snapshot of everything, checked it, removed Rune Runtime from the server, shut OpenClaw down and installed Hermes the same morning. The file that says who Rune is links straight into Aetherkeep, so the first session on Hermes was already Rune. From May to July it took 760 tasks from kanban boards, handed them to specialist agents and reported back over Telegram.
481 of those 760 tasks were marked done. A task can take several runs: one run is one attempt.
On the busiest board, 3,309 of 3,583 runs crashed, and 3,193 of them were one task. It asked for a skill that wasn't installed on the server, crashed within a second, and every time Hermes's safety cut-off gave up on it, the scheduler put it straight back in the queue. I traced the loop, installed the missing skills and checked that all 62 active tasks could load theirs.
What it taught me
- 1
Logs the agent can't reach
Rune's own log of 18 February was deleted with everything else, so the one record that could say what happened is gone.
- 2
Off means off
On 28 February one agent tried to stop the other, and systemd restarted it within seconds.
- 3
Unclear work goes to a person who hears about it
Novaegis's proposals waited with no one acting on them, and later its alerts failed without anyone noticing.
That is why the agents I build for others are logged, easy to switch off, and hand anything unclear to a person.
Today Rune runs on Hermes Agent, on my server and my own PC, with its memory still in Aetherkeep.