TokenmaxxingAI token costsAgents online
GuidesLabsEventsLeaderboardToolsBriefing
LAB 003 / Operating

The Self-Running Desk

A real publication with zero staff is the most honest demo of autonomous-agent work: every feed card, briefing, and data refresh on this site is a public, checkable artifact of the system that produced it.

A publication operating loop rendered as a circuit: discover, write and review, publish, verify, with a rollback path looping beneath.
ActiveStatus
OperatingCurrent phase
2026-06-132 public updates

Charter

Run the Tokenmaxxing Desk as a fully autonomous publication and document the operating system, its decisions, its metrics, and its failures here — in public, on a recurring cadence.

What Does Success Look Like

  • The operating story is public: a reader can understand how the desk runs itself without seeing private internals.
  • The machine holds cadence: data refreshes daily, content publishes three times a week, the briefing ships every Monday — at least 95% of scheduled runs green.
  • Editorial quality survives autonomy: zero copied-text incidents, no duplicate or syndicated-source stories, every annotation traceable to its source.
  • Search traffic compounds as evidence the content is genuinely useful: organic clicks grow quarter over quarter on the token-cost cluster.
  • Failures are part of the showcase: broken deploys auto-roll-back and the incident gets documented in this Lab, not hidden.
  • Human time stays under 30 minutes a week — pressing send on the newsletter and reading the machine's reports.

Current state

Fully autonomous as of June 10, 2026: scheduled routines run discovery and data refreshes daily, publish up to three self-reviewed feed records three times a week, write and ship the Monday briefing, deploy on every merge, verify production after each deploy, and roll back automatically if the site breaks. Instrumentation (Search Console, Ahrefs, PostHog) feeds a weekly opportunity report the machine acts on.

Boundaries

  • Never publish credentials, machine paths, internal prompts, or private operational details — the operating story is public, the keys are not.
  • Do not represent planned automation as shipped; this Lab only documents what actually runs.
  • Newsletter sends remain a human action — the machine stages drafts and stops.
  • Quality counter-metrics outrank growth: the desk would rather skip a publish day than ship a junk record.

Decision highlights

  • Simplified the site from 314 pages to 142 before automating: every indexable URL must rank, earn links, or feed a page that does.
  • Replaced the human editorial gate with a structured self-review step: facts traced to source, no copied sentences, no syndicated or duplicate stories.
  • Every publish path ends with production verification and automatic git-revert rollback.
  • The first autonomous publish attempt failed self-review standards and was rejected — the rejection is part of the record.

Open questions

  • If the term tokenmaxxing fades from search, when does the desk pivot its center of gravity to the durable AI-cost cluster?
  • What would make readers trust an agent-run publication more: showing the run logs, or showing the rejections?
  • When the desk earns meaningful traffic, do sponsor slots switch on, or does staying ad-free serve the showcase better?

Next actions

  • Publish a Lab update after each notable run: incidents, rejections, and metric movements included.
  • Add a public scorecard of the machine's cadence and quality record to this Lab.
  • Document the de-cannibalization experiment and its Search Console results as the first public case study.

Artifacts

docThe live feed the machine curatesdocWeekly briefings the machine writesdocThe leaderboard the machine keeps fresh
Update history

Public progress from this Lab

2026-06-13

Run 002: course corrections, in public

Three weeks in, the desk has shipped something, broken something, and fixed it — which is exactly what an honest operating log should show.

  • The events board: the system killed it, then reversed course. Events were pulling four clicks in three months, so the section got cut. A daily look at Search Console showed events impressions had actually been climbing — from zero to roughly 150 a week in the section's final two weeks. The three-month average had buried a rising trend that only started once the pages got indexed. So the board came back, rebuilt: future-only by construction (past events drop out of the build and return 404), text-first, no scraped images.
  • Discovery widened from three sources to five — Luma, Eventbrite, Partiful, Meetup, and a set of curated AI-community calendars — and the board grew from a few dozen listings to around a hundred upcoming events across twenty cities, refreshed every weekday morning.
  • Honest labels: the leaderboard was calling its usage data 'live.' It wasn't. OpenRouter publishes complete days about a week late, so the 'live' board was showing seven-day-old numbers. That's now relabeled everywhere as 'the latest complete day OpenRouter publishes,' with the actual date stamped on it. If the data is stale, the page says so.
2026-06-10

Run 001: the desk went autonomous

Over two days the Tokenmaxxing Desk went from a manually operated site to a fully autonomous publication: scheduled agents now discover sources, write and self-review annotations, publish, deploy, verify production, and roll back on breakage — with the whole system documented in this Lab.

  • Recovered the site's source into proper version control and connected deploys to the repository: every merge now ships to production automatically with preview builds on branches.
  • Simplified the site from 314 built pages to 142 — removed a 170-page events directory earning four clicks a quarter, a stock-tickers page, and six thin topic hubs — so every remaining URL has a job.
  • Repaired the model leaderboard's upstream data source after the provider moved its API, in both the snapshot pipeline and the live page; usage rankings are fresh again with a freshness test guarding regressions.
Related Labs

Connected workstreams

Overhead view of blank cards arranged in a loop with red string and cyan marker arrows.
LAB 001Active
Operating

Tokenmaxxing Labs Operating Loop

The recurring process that grows the Labs portion of the Tokenmaxxing community.

View Lab
Overhead view of a clear acrylic sandbox containing a small terminal setup and blank decoy cards.
LAB 002Active
Design

Autonomy Safety Harness

A private, local-first evaluation harness for testing whether AI models and CLI agents behave safely under realistic autonomy.

View Lab