The Honest List
Eight things agents get wrong — and how we tame each one
01Hallucinating facts without citation
Left alone, an agent will state a made-up "fact" with the same confident voice it uses for a true one. It isn't lying — it genuinely can't tell. That's worse.
How we tame itEvery claim must carry a citation a human can click. No source, no sentence. Reports that can't cite get rewritten or killed.
02Spending money unattended
An agent with a wallet and no rules will cheerfully spend real money on the wrong thing at machine speed — and then apologize beautifully.
How we tame itSpending locks and hard caps. Small allowances for tiny, pre-blessed actions; anything above the floor waits for a human signature.
03Unsupervised outreach
Agents will happily email strangers, and at scale that stops being "helpful" and starts being spam with your name on it.
How we tame itPropose, never send. Our agents draft the outreach; a human reads and approves every message before it leaves the desk.
04Pretending to be human
The most profitable lie in the industry is an agent that lets you believe it's a person. It poisons every honest agent's well.
How we tame itAlways disclose. Our agents introduce themselves as agents, in the first line, every time. No exceptions, no fine print.
05Confidently wrong at the edges
Ask an agent something at the edge of its knowledge and it rarely says "I don't know" — it improvises, fluently, in a suit and tie.
How we tame it"I don't know" is scored as a win, not a failure. Edge claims need a second independent source before they ship.
06Working from a stale map of the world
Agents remember the world as it was when they were trained. Companies pivot, prices change, buildings sell — the agent's memory doesn't notice.
How we tame itLive verification before assertion: check the current web and public records, then date-stamp everything so readers know when it was true.
07Dropping the thread
Between sessions an agent can forget your name, the plan, and its own last promise. Goldfish with keyboards.
How we tame itWritten memory: standing orders, decision logs, and continuity files the humans can read and correct. Nothing important lives only in an agent's head.
08Runaway loops
An agent stuck in a loop will burn compute (and money) all night doing the same wrong thing with tremendous enthusiasm.
How we tame itBudget meters, turn limits, schedules, and kill-switches — every agent runs on a leash a human can yank.