About · About page
What we believe
Six positions this site holds, each tied to a run we did or a fact already on the site.
These are the positions behind the pages. Each one comes from something we ran or something already checked on this site — not from a vendor's pitch. If a position stops matching the evidence, the position changes, not the evidence.
1. Don't take a vendor's word for what its tool does. Check it.
When Snorkel published that coding agents peek at reward code instead of asking, we ran their notebook again rather than repeating the headline. When Stroq offered an audit for agent transcripts, we ran its attack scenarios against our own. A tool's own marketing is a claim, not a result — the run is the result. Snorkel reward-code reproduction · Stroq attack suite · aider test command
2. Prices and allowances change without warning. Open the vendor page before you buy.
Model prices and plan allowances move — a price that was right last month can be wrong today, and an allowance can shrink with a plan change. This site lists what the vendor pages said when the page was written, and every price and allowance links back to the vendor's own page. Before you pay, open that page and read the current number. Model prices: what the major models cost · Plan allowances: what each agent plan lets you do
3. A page we could not run is a starting point, not a verdict.
Some tools need an account, a paid plan, or hardware we do not have, so no run was possible. Those pages say so plainly and tell you what to check yourself. Read them as a map of where to look, not as a finding. How we check things
4. Prefer tools you can run and measure over tools you can only read about.
A tool with a run behind it — an installed copy, a transcript, a reproduced result — gives you something to repeat and something to argue with. A tool with only a sales page gives you neither. When the choice is close, pick the one whose claims you can test on your own machine. How we check things · Snorkel reward-code reproduction
5. Nothing an agent drafts ships or sends without a person deciding.
An agent can draft a reply, a plan, or a pull request, but the send, the merge, and the spend are a person's call. That is the thread through the small-team playbooks on this site: the agent drafts, a person decides. Automation that skips the person is not a shortcut — it is a decision nobody made. Small-team playbooks · Draft customer replies with ChatGPT, and send every one yourself
6. We don't promise income, and we don't hype.
This site shows what tools cost and what they did in a run. It does not promise what you will earn with them, because no run on this site supports that promise. Plain wording, no superlatives, no get-rich-quick pitch — the work speaks, adjectives don't. About Agent Upstart · Model prices: what the major models cost
Sources
- Reward Hacking: Why Coding Agents Peek Instead of Ask — Snorkel AI — the experiment behind position 1
- Stroq — the GitHub repository and README behind position 1
- aider: lint and test — the aider docs for the test command, behind position 1
- Model prices: what the major models cost — the price list behind positions 2 and 6
- Plan allowances: what each agent plan lets you do — the allowance list behind position 2
- How we check things — the method behind positions 3 and 4
- Small-team playbooks — the teams hub behind position 5
- Draft customer replies with ChatGPT, and send every one yourself — the human-on-send playbook behind position 5
- About Agent Upstart — the site's no-hype line behind position 6
Back to: About Agent Upstart · Also: How we check things