One popular claim about building software with AI agents per issue, put up against the evidence. Most don't survive. The ones that do get a fair hearing.
Don't worry, the agent asks before it runs anything. A human approves every command.
The human approves the command; the damage is in what the command sets off. Sandboxes and effect checks keep you safe. The click keeps you busy.
If the AI makes up a package, the install just fails. No harm done.
A made-up package fails to install only until somebody registers it, and the AI will tell anyone who asks which name to register. It's a wrong number that a stranger can go and buy.
Agents score 99% on SWE-bench now. Coding is a solved problem.
A benchmark gets solved. A job doesn't. Ninety-nine percent on a cleaned exam tells you the student can pass that exam. It tells you nothing about the shift at a real hospital.
Ninety percent of agent PRs get merged. Agent code is production-ready now.
Ninety percent of agent PRs got approved. Nobody has measured how many were right. A merge is a signature, not a test result.
Last updated . Reuse with credit under CC BY 4.0: “The Dabbawala Protocol, thedabbawalaprotocol.com”.