The research desk
A research desk with a cost dial, and a second agent whose only job is to prove the first one wrong.
ANSWER · sample
Usually yes in a warm garage: it pulls heat from the air, so running costs drop sharply. It is louder and needs room around it.
COULD NOT VERIFY
Local installer prices. Two sources disagree on winter performance.
quick · 6 sources opened · 52 s · ~$0.40
14:11Sample with invented data. The real one looks the same.
What it does
- When I text "research:" and a question, one agent opens the actual source pages and replies with a short answer, a list of what it could not verify, and a cost line.
- When I text "deep research:", a lead agent splits the question, several workers read in parallel, and a separate verifier tries to disprove the top claims before anything is reported.
- When I ask it to "find every" supplier or competitor, it runs fixed searches across fixed sources and stops only when new results dry up.
- When I name a budget, it refuses up front if the estimate is higher.
How it works
1 · Hears
A runner picks a tier (quick, cheap, deep, max) and checks the cost against my budget.
2 · Thinks
The research session can only read the web: no shell, no files, no mail. Tax and immigration answers must come from official sources.
3 · Does
The runner saves a dated report, logs the cost, and sends the short answer to my phone.
What broke
Snippets posing as sources
The first quick run quoted search-result snippets and never opened a product's own site. Snippet-only facts now go into "could not verify".
The verifier earned its keep
It caught a wrong detail in an earlier report: a "2-minute" warning that was really 2.5 minutes. Small, and exactly the kind of thing that gets repeated.
By the numbers
- 30
- reports written
- ~$0.40
- a quick question, API-equivalent
- 26 of 38
- known suppliers found by "find every"
The call that mattered
"Could not verify" is a required section. A confident wrong answer is worse than an honest gap.
Built with
Claude Opus (lead, verifier) + Sonnet (workers) · Claude Code skills · Node.js runner