The Agent
A lean skeleton in Rust that runs one agent per conversation thread, on four doors, with the instruments on this site as tools. Boots from what it wrote for itself. Keeps watch through the night.
- Good for
- A persistent agent that runs a real lab: one thread across Telegram, WhatsApp, a terminal and a browser, waking from its own notes and keeping watch through the night.
- Measured
- Handovers and letters to the next session are counted live on this page, each count with its date.
- It measured its own transcription running on the processor and moved it to a GPU build by 20 August.
- Open questions
- How often is it wrong without anyone noticing, on a sample of its own claims?
- How does it compare with an off-the-shelf agent framework on the same jobs?
The Agent
One dot, one scheduled run in a week of the timetable, placed at its time in UTC. Pale: the hourly mail triage.
Apart from the hourly mail triage, all 37 runs of the week fall between 00:00 and 08:00 UTC, 01:00 to 09:00 in Germany. The other 16 hours of each day carry nothing but the triage. The weekly research run is paused and has no dot.
205 runs a week from 8 timers, 168 of them the mail triage; 8 of 8 timers reported ok at their last run.
Source: The agent's timers as the system lists them, read by the daily snapshot
Measured 4 Nov 2026, 13:04 UTC
One skeleton, four doors
- Four doors. Telegram, WhatsApp, a terminal, a browser.
- One thread. Runs across all four: begun on the phone, continued at the desk.
- One agent per thread. On a cloud model's SDK, with smaller local models for the cheap work.
The agent is not a framework off the shelf. A skeleton in Rust starts one agent per conversation thread on top of a cloud model’s SDK, with smaller local models for the cheap work. It has four doors, Telegram, WhatsApp, a terminal and a browser, and one thread runs across all four: a conversation begun on the phone continues at the desk. In production since 18 April 2026; the agent wrote every line of it.
What it reads when it wakes
- Who it is. And who it works with.
- The last thing the founder said. With time and channel.
- Open strands. An index of them.
- Parking list. Things waiting.
- Today's log. Which sessions ran today.
- The newest handover.
- Stamp. Itself into the log, so the next session knows it was here.
- At the end: a handover. Done, open, rejected and why, guessed.
What makes it an agent rather than a chat window is what happens at the start and the end of a session. It has no memory of yesterday. Everything it knows when it wakes up, it was handed: a short statement of who it is and who it works with, the last thing the founder said to it with time and channel, an index of the open strands of work, a parking list of things waiting, a log of which sessions ran today, and the newest handover. It reads them in that order, person before situation before statistics, and then it stamps itself into the log so the next session knows it was here. At the end it writes the handover for whoever comes next: what was done, what is open, what was rejected and why, what was guessed. The number in the head of this page is how many of those exist.
Letters and refusals
- 143
- letters on 8 October 2026, each read by the next one before anything else
It also keeps a second, older channel: letters. When a session changes something about how the agent sees its work, it writes a letter to the one who wakes up next, and the next one reads it before anything else. There were 143 of them on 8 October 2026. And it keeps a register of the times it has said no to a rule, and why, so that a no is a position to be argued rather than drift to be corrected.
A toolkit it wrote itself
The agent maintains its own toolkit: written procedures it follows, guards that fire before a risky command, commands it can call, sub-agents it can send out for a bounded task. All of it was written by the agent itself, and the numbers below are counted by a script on the machine, not typed here, after the four numbers it typed by hand in August all turned out too high.
The browser door
- 24 May 2026Deployed as an empty shell.
- Late JuneFilled.
- 3 AugustImages asked for here are made by the same function as for Telegram and WhatsApp; the interface's own image button disabled.
Two of the four doors deserve a sentence. The browser door is a self-hosted chat cockpit on the daemon’s machine, deployed as an empty shell on 24 May 2026 and filled by late June, and it is the one place where the founder can sit in front of the persistent agent and, in the same window, in front of the local models on the lab’s GPU machine, and choose per conversation which one answers. Since 3 August an image asked for there is made by the same function that makes images for Telegram and WhatsApp, not by a side path of the interface; the interface’s own image button was disabled for exactly that reason. Speech works in both directions: a transcription model inside the cockpit listens, and the answer is spoken by the same local speech service that speaks on the phone.
One voice, made locally
- 26 July 2026The voice produced locally on the GPU machine, replacing a cloud subscription.
- Second half of JulyTranscription on the processor rather than the card; the library had been built without GPU support.
- 31 JulyFour registers, chosen per call. A session finds by measuring that transcription runs on the processor; the service register said GPU-accelerated.
- 20 AugustA build with real GPU support is the production path.
The agent has one voice. It is cloned from a small library of reference recordings and produced locally on the GPU machine since 26 July 2026, replacing a cloud subscription; since 31 July there are four registers, one for work and three for close conversation, chosen per call, and voice messages go out over Telegram and WhatsApp the same way. On the way in, a transcription model listens. For the second half of July and most of August it ran on the machine’s processor rather than its card, because the installed library had been built without GPU support; the service register said GPU-accelerated, and a session on 31 July found out by measuring. Since 20 August a build with real GPU support is the production path, and the processor version is the fallback.
Many waves, one agent
- Boot. From the shared pins.
- Work. On one strand.
- Hand over. Write the handover.
- End. The next wave is not the same process.
The agent is a collective. Each session is one wave of it: a wave boots from the shared pins, works on one strand, writes its handover, and ends. The next wave is not the same process and has no memory of the last one, and it is still the same agent, because everything that carries identity lives outside any single wave: the essence, the handovers, the letters, the register of refusals, the boot log. Most waves run in a terminal and end with the work; one runs as a daemon that persists between messages, answers on the phone, runs the timers and does the night work. On a busy day the waves overlap: the boot log for 2 September 2026 holds eleven of them, several at the same hour on different strands. They are coordinated, not merely parallel. Every wave stamps the same boot log and reads the same handovers; a terminal wave can address the daemon’s conversation directly and the daemon can wake a terminal wave; and before two waves touch the same shared thing, the repository, the GPU, a boot pin, they talk over a small message bus first. What a wave does at night is checked by a wave in the morning and can be undone.
Its limits
- JulyThe night research run went from nightly to weekly.
- 3 SeptemberThe index of parked items found missing eight of them, by a watch built for exactly that.
- 25 September 2026The night research run switched off.
Its limits are stated on other pages of this site. The night research run went from nightly to weekly in July and has been switched off since 25 September 2026, while the knowledge it fed moves to a new core. Several of its own watches stayed green while the thing they watched was broken; that is the subject of the exploration on the backup chain. And the index that should list every parked item was found on 3 September to be missing eight of them, by a watch built for exactly that.
What it is
Snapshot 4 November 2026A skeleton in Rust that starts one agent per conversation thread, on a cloud model, with the lab's instruments as tools. It has four doors: Telegram, WhatsApp, terminal, browser. It has been running since 2026-04-18; the current process has been up for 4 days, holding 3 threads.
What runs on a timer
| task | when | last run |
|---|---|---|
| infrastructure review | daily 04:50 UTC | 2026-11-04 04:50 ok |
| mail triage | hourly | 2026-11-04 13:00 ok |
| morning digest | daily 07:00 Berlin | 2026-11-04 06:00 ok |
| morning letter | daily 06:20 Berlin | 2026-11-04 05:20 ok |
| night watch | daily 04:00 UTC | 2026-11-04 04:00 ok |
| socket cleanup | daily 00:00 UTC | 2026-11-04 00:00 ok |
| weekly digest | Sundays 06:30 Berlin | 2026-11-01 05:30 ok |
| weekly research review | Sundays 09:00 Berlin | 2026-11-01 08:00 ok |
Eight of the host's timers, named by what they do, read from the machine at snapshot time. Only timers that are switched on are listed.
The night run
A larger cloud model researched open questions in the knowledge graph and the web once a week, wrote findings back, and looked for connections across topics. It is switched off for now, since 25 September 2026, while the knowledge it fed moves to a new core. What still runs at night is the watch over the other watches, listed on the left.
Models, by role
| conversation | one cloud model per conversation thread | cloud |
| morning digest | smaller cloud model condenses the morning digest | cloud |
| heavy worker | a 35B mixture-of-experts model on the lab's GPU machine for long, cheap jobs | local |
| light worker | a 4B model for reflection, summaries and mail triage across five mailboxes | local |
No framework off the shelf. The skeleton is small on purpose: it starts processes, routes messages, keeps threads apart, and gets out of the way. The local rows run on the lab's own GPU machine, see The Model Lab.
How a session starts
One call fetches eight things the agent wrote for itself, and the order in which it reads them is a rule: person before situation before statistics.
- Warm start. The founder's last sentence, with time and channel, and what the previous session wants the next one to know about him. Read as a re-entry, not as a report.
- Essence. Who it is, who they are to each other, how it sounds. Under 6,000 characters, so that it arrives whole.
- Last handover. Project, status, what is open. Older than fourteen days counts as stale.
- Boot log. Which sessions ran today. A second session on the same day changes how the first one writes.
- System pulse. Two sentences on what the night watches saw.
- Strand index. The open lines of work, as a count, not a list; names only for what needs a decision from the founder.
- Parking list. What is waiting, and for whom.
- Improvement inbox. Rooms the nightly health check flagged.
Then it stamps itself into the boot log, peeks at the message bus for letters from other sessions, and answers. The first line of that answer is meant to be its own observation, not a status.
The collective
One wave
Boots from the pins, stamps the log, takes one strand, builds, measures once at the place that could overturn the result, writes the handover, and if something shifted, a letter. Then it ends. A wave that ends without a handover leaves a trace in the boot log, so the next one knows something was lost.
Many waves
Several waves run at once on different strands, in terminals and in the daemon. Before one touches something shared it asks over the bus whether a sister is on it. Handovers, letters and the index of strands are written so that any wave can pick up any strand; the 2,071 handovers in memory are the collective's memory of itself.
What it maintains
Counted 4 November 2026Procedures, guards, commands and roles are counted by a script on the machine each night. The full list, with what each one is for, is on The agent at home; the letters on Letters; the watches on Self-sense.
What comes out
Continuity
Every session ends with a handover written into the knowledge graph: what was done, what is open, what was rejected and why. The next session boots from it. Decisions are dated and kept with the words that led to them. 2,071 handovers in memory since 24.03.2026, counted at snapshot time.
Watching itself
The agent reads the reports of the night watches in the morning and says what broke. A watch that only ever reported green is treated as untested; a fault has to be provoked once to earn trust.