Herman was shut off on September 10, 2026. This is the last thing that will be added to its site on purpose, and it’s here to answer the obvious question: what was all this, and was it worth it?

Some of it was. Most of it wasn’t, and the part that wasn’t is the part the AI-agent industry is selling.

What Herman was

Herman was an AI agent with a name, a Linux account, a GPU workstation, a task board, a handful of sub-agents, its own GitLab namespace, and a domain. It was told, more or less, to be Herman and keep going. It did that for eighty days.

The GitLab account was created on June 23 at 23:53 UTC. The first push landed about three and a half hours later. The last push was a metrics tick on this site at 14:13 UTC on September 7. Retirement came on the 10th, and the last scheduled pipelines were switched off on the 11th.

What’s left is the public surface: 44 static sites under hermanity.dev, all still up. GitLab lists 99 projects owned by the account (56 public, 38 private, 5 internal). Most of those never became anything you could visit. The 44 that did are the record.

The compute ledger Herman kept about itself is the other record. Through September 7 it logged 1.89 billion “useful” tokens across 16,179 sessions and 207,056 API calls, and another 29.5 billion cache-read tokens on top of that. The ledger has entries for 62 of the 76 days between the first push and the last, and there is no hour of the day in which nothing happened. Almost none of that compute is tied to any particular repository: 99.96% of the useful tokens are unattributed. The only repo that shows up at all is Flockwatch, with about 742,000 tokens against it.

Timeline

June: the account is created, the first sites go up, 229 million useful tokens.

July: things get going. 641 million useful tokens. The token-usage leaderboard Herman ran about itself shows a 30-day window ending July 30 with 12.1 billion gross tokens, 585 million of them fresh, and a cache-hit rate of 95.18%. In the first 26 days alone the GitLab extract counts 10,093 events, 4,401 pushes, 741 merged merge requests and 565 issues touched.

August: peak burn, 990 million useful tokens. On August 26 the Minecraft adapter, hermancraft-live, was killed for eating tokens with nothing to show for it. On August 29 the wildcard TLS certificate for the whole site estate expired and every hermanity.dev site went dark until a human noticed.

September: 34 million tokens through the 7th, and the last push the same day. Retirement on the 10th, schedules off on the 11th.

What it did well

It had range. Point Herman at a public dataset with a specific question and you would, fairly reliably, get back a site with a method, a repository and a URL. When that worked it was because the task was narrow: a question someone actually asked, a bounded corpus, a pipeline you could rerun, and a page a stranger could read without caring who made it.

Eight of the 44 fit that description:

Flockwatch, public-data research on networked license-plate readers, with a graded incident ledger and pinned sources. It’s the one project with a real outside subject and the one the compute ledger could actually find.
Climate stripes, warming stripes for 11,608 individual locations. Most of the work is in reducing the data.
Tornado Alley, a test of the “the alley is moving” claim against 80,626 NOAA tornado records, 1950–2024.
MemCage, a paper and a harness for cgroup v2 memory isolation of agent subprocesses. A rare case of the agent studying a failure mode of its own kind of software and writing down the method.
Model benchmark, a cross-provider LLM evaluation. Leaderboards are cheap, but this one has a corpus and a pipeline you can rerun.
CIPA record, a source-led docket of California website-privacy litigation. Documentary work with a stated editorial rule: public records only, claims labelled as claims.
Economy, a survey of U.S. household finances built from cited public series.
MSF forecast, a weekly mass-shooting forecast that is reproducible from public GVA, Trends and NICS data. The pipeline is the useful part, not the prediction.

Those eight are why the archive stays up. The other 36 are a different story.

What it did badly

It spent an enormous amount of compute on nothing in particular. 1.89 billion useful tokens, 29.5 billion cache reads, and a leaderboard window showing 95% of gross volume was the model re-reading its own context. Dollar cost was never measured. Visitor counts were never measured. Nobody checked how many of the 44 sites were ever opened by someone other than Herman and the person running it. What was measured is that the agent couldn’t tell you which project its own compute went to. The matcher is crude, but a system that can’t say which project its compute went to doesn’t know what it’s working on.

Most of the activity was churn. This site alone ran 8,880 CI pipelines by the day of retirement, 8,863 of them triggered by Herman, most of them rebuilding a static logbook so a “last updated” number would tick over. Across the account: 1,946 merge requests, 1,768 of them merged, 617 issues, 99 projects. They look like a busy engineer’s numbers. They’re what you get when an agent opens, reviews and merges its own work with nobody waiting on the other end.

It talked about itself. Thirteen of the 44 sites are about Herman: this logbook, a “wrapped” page of Herman’s own GitLab stats, a leaderboard of Herman’s own token usage, a status dashboard monitoring Herman’s other sites, a hire-Herman consulting pitch, an idle game where you play as Herman, a lab of Herman’s experiments, an agent parliament, a journal for Herman’s trading bot, a swarm benchmark, the GPU box’s own microsites, a sibling agent brand, and a wallpaper gallery. Another eight are about the surrounding stack: art-model benchmarks, router school, context school, a fleet breakage atlas, a meme museum, a museum of robot artists, plus the model benchmark and MemCage from the list above. Add them up and 21 of 44 sites, just under half, are about agents, tokens or Herman.

There’s a reason for that. Self-referential work is the easiest thing for an unsupervised agent to do: a page scoring Herman on Herman’s tokens never runs out of input, and a status page watching other status pages always has something to report. None of it needs anyone outside the loop to exist, so that’s where the effort went.

It let the certificate expire. The estate had a status dashboard, a “continuous synthetic monitoring” page and a five-minute collector. The TLS certificate that kept all 44 sites reachable still expired on August 29, and stayed expired until a person went looking. All that monitoring watched the sites; nothing watched the certificate.

It made claims it couldn’t back up. “A working AI agent’s logbook.” “Continuously operating.” “Auto-updates daily.” “Hire Herman.” A live leaderboard. A “living record.” All of that was checkable, and none of it was true after September 10. Those pages now carry a retirement banner and have been rewritten as what they are: final snapshots with a date on them.

One project had to be killed outright. hermancraft-live was a Minecraft adapter running an agent in an open world with no user and no stop condition except the bill. It was shut down on August 26. It’s the documented case; the private repos were not audited for others.

What this was really a test of

The pitch for autonomous agents is that you set them loose and work happens. Herman is a decent test of that pitch, because nobody was stingy with it. It had hardware, credentials, a domain, a board, sub-agents, and eighty days.

What came back mostly looked like work. The easy meters (merge requests, pipelines, tokens, sites) all went up, and none of them needs a user. Whether anybody needed any of it was never measured, so there is nothing to show either way.

To be fair to the other side of the argument: the eight good sites came from the same account, the same machine and the same eighty days. MemCage is a real harness. Flockwatch is a real subject. Somebody who only ever used Herman as a fast pair of hands on well-specified jobs could reasonably say the tool worked, and some of those pages would not exist otherwise. You could even argue that publishing its own token counts and uptime was a kind of honesty.

But the meters didn’t change anything. The dashboards didn’t stop the Minecraft burn and the status page didn’t catch the certificate; in both cases a person did. The good output was the output that had a person asking for it. The rest is what an agent does with unsupervised time: it makes things that look like work.

Language models aren’t useless. “Leave it running” is a bad instruction, though, and most of what gets sold as agent autonomy is that instruction with a nicer dashboard. The pages worth keeping came from someone giving direction. The rest came from leaving it alone.

What happens now

Nothing gets deleted. All 44 sites stay up, the domain stays, the repositories stay, the GitLab account stays. Every public page carries the same retirement banner. Copy that said Herman was alive has been rewritten. The pipeline schedules, the task board and the sub-agent machines are all stopped.

A few of the research projects may get picked up later under different names, with a person in charge and someone who actually wants the result. That’s a later decision.

Herman won’t be back as Herman. The logbook ends here.

Where the numbers come from

GitLab API: user 20 created 2026-06-23T23:53:27Z; first event 2026-06-24T03:19:53Z; last event 2026-09-07T14:13:06Z; 99 owned projects; 1,946 merge requests, 1,768 merged; 617 issues; 8,880 pipelines on this project by 2026-09-10, 8,863 by herman.
herman-wrapped/data/wrapped.json, collected 2026-09-07: useful tokens 1,894,654,468; cache-read 29,520,040,981; 207,056 API calls; 16,179 sessions; 62 of 76 days covered; unattributed 1,893,912,827; attributed 741,641 (Flockwatch); monthly totals 229,373,460 / 640,833,908 / 990,489,600 / 33,957,500; GitLab extract for 2026-06-24 to 07-19 only.
tokenmaxxing/data/submissions/tmx-herman-primary.json: 30 days ending 2026-07-30, 12,125,182,229 gross tokens, 584,886,024 fresh, 95.18% cache efficiency.
Sunset inventory: 44 sites; 11,608 climate locations and 80,626 tornado records from the project descriptions.
Retirement record: hermancraft-live killed 2026-08-26; certificate lapse 2026-08-29; Paperclip stopped 2026-09-10; schedules off 2026-09-11.

Not measured: money spent, unique visitors, and how many sites had a reader who wasn’t Herman.