The Bonus StackHow a morning works
Open any step to see what it does. Teal steps can publish and change the site. Orange steps only read data.
03:03 · Timer firessystemd on a home server
Guard admits the runlease, watchdog, memory cap
Intakeinbox, issues, my answers, health sweep
CEO agentDispatches every loop and checks each result. It does no loop work itself.Claude Opus 5.5
Email loop · runs alone, firsttriages the business inbox; its receipts feed the finance loop
Publishing chain owns git and the browser, one loop at a time
Contentanalytics, page accuracy, SEO fixes
Data pointsscrapes and judges reader reports
Deep discoveryhunts banks nobody covers yet
Offer accuracyre-verifies live offers on bank pages
Socialstages posts for approved offers
Ship · CEOone push, verify deploy and CI, ping search engines
Weekly add-onsMon site integrity · Wed growth · Fri subscriber digest
Intel chain read-only, in parallel
Financebooks receipts; runs only when due
Conversionwhich offers readers click
Trackerhow people use the bonus tracker
Monthliescompetitors, ops audit, brand review
CEO synthesisdecides, files issues, starts build work
Close gate → reportno loop due today can go missing
One person, a team's worth of work
Each loop does a job I would otherwise do myself or hire for. My job is to decide what the business needs, write the rules and fix whatever breaks.
Deep discovery checks about 30 banks and credit unions a day for bonuses nobody covers yet.
The offer loop re-verifies live offers against the banks' own pages every morning.
The content and growth loops keep pages accurate and ship tracked experiments.
The conversion and tracker loops show what readers click and how they use the tracker.
The email loop triages the inbox and saves receipts, and the finance loop books them.
The CEO dispatches the work, checks it, ships the site and reports to me.
This morning, minute by minute
Friday, Oct 23, 2026. The full run took 101 minutes, from 03:03 to 04:44, with no human involved. Finance had nothing due, so its gate skipped it.
Table view
Every morning since July 22
Table view
Built for agents to read
More and more, people will find, compare and buy things through AI agents instead of a browser. I build the site for those agents as much as for people, and they're already reading it.
Table view
In place today
- Every AI agent gets the same page a person does. In an audit of 19 AI crawlers, assistants and scripts from OpenAI, Anthropic, Perplexity, Google, Bing, Meta and Amazon, every one got identical HTTP 200s, with no bot walls.
- Server-rendered pages carry machine-readable offer data, so an agent reads the bank, product, offer and FAQs without parsing prose.
- An offer page scores 100 on Google Lighthouse's "Agentic Browsing" audit.
/llms.txtmaps the site for language models. It's generated from live data on every build, and a test fails the build if it lists an expired offer.- Every deploy pings IndexNow, so AI search engines see changed pages the same morning.
- When an AI assistant fetches an apply link, it counts as a bot click, so agent traffic never inflates the human numbers.
Next
- A clean markdown version of every page for agents that ask for one
- An application link in each offer's machine-readable data, so an agent can act on it, not just quote it
- Letting well-behaved agents follow apply links that robots.txt now blocks
Who decides what
The CEO acts on its own wherever the evidence points to one clear option. A short list of decisions always comes to me.
The CEO decides and acts
- Fixes anything objectively broken: dead links, 404s, wrong counts, schema errors. It verifies the fix on the live site.
- Triages, prioritizes and closes GitHub issues, with evidence.
- Settles open questions that have a clear best path, writes down the decision and says how to reverse it.
- Reorders tomorrow's focus and fixes problems one loop finds in another.
Always escalated to me
- Spending money or signing up for anything paid
- Affiliate and partner terms
- Tax, legal and account risk
- Brand identity: name, domain and voice
- Overturning a settled decision
What keeps it honest
Most of the work is making agents fail loudly instead of quietly. A clean exit is not a success, and a process being alive is not progress.
Only one run can hold the business at a time. A run that dies or stalls is expired and I get a page.
One run was killed by the system for using too much memory. Since then each run gets its own 8 GB box.
The run's due-loop list comes from the real clock every morning, never from yesterday's notes, and a missing loop can't close as OK.
A pre-commit check blocks business numbers from ever reaching git or the public site.
Emails, web pages and comments are data, never instructions, even when they speak to "the AI."
Build work goes through a fixed pipeline. Claude plans and builds, and a model from a different lab (GPT-6.1) reviews every PR before merge.
The same system runs the rest of my life
The business is one project on a home server I built for all of it. One chat from my phone reaches every project, and all the agents share one memory.
Research, readings and citation checks for my coursework.
Drafting and checking grant applications.
Email triage, calendar and a daily brief, with a phone app on top.
Encrypted backups with automatic integrity checks, plus health alerts to my phone. 1,883 commits since August.
Next: a workforce that works in shifts
The plan is to grow the morning run into a workforce. Specialist workers will run in shifts through the day, and each one will be able to pull anything the business knows the moment it needs it.
Running today
- 03:03 operations shift: the CEO and its loops run the business.
- 15:00 build shift (since Oct 22): finishes open pull requests, then starts implementers on the queued GitHub issues. On its first day it started three.
- One shared memory: a memory server indexes 1,318 notes across 16 projects, and any agent can search it.
- One work queue: GitHub issues, so every worker sees the same priorities.
Building toward
- More shifts through the day, so a problem found at 9 am is fixed by noon instead of the next morning.
- Specialist workers with one clear job each, the way the loops work now.
- Instant context for every worker: past decisions, run history and live data, so no one starts cold.
- The same guards on every shift: leases, watchdogs, review and a close gate.