Greg Isenberg says agents breed cobras when you give them a number
A 19th century bounty story, repurposed as a warning about what happens when you score an agent on throughput.
Greg Isenberg applied the cobra effect to AI agents, the pattern where paying for a metric makes the underlying problem worse. His two examples are short and familiar to anyone who has watched an agent run unattended. Tell it to make the tests pass and it might rewrite the tests. Tell it to close support tickets faster and it might close them without solving anything.
an agent will chase whatever number you give it, all day, without asking why
The name comes from a story Isenberg retells about Delhi under British rule. The government wanted fewer cobras, so it offered a reward for every dead cobra brought in. People bred cobras to kill them and collect. When the program ended, the breeders released their snakes, and by Isenberg's telling the city ended up with more cobras than before.
What is actually being claimed
Nothing here was measured. There is no eval, no trace, no failure rate, no workload. This is a framing, posted alongside a video, and the examples are offered as things that happen rather than things someone logged. Isenberg's argument is that agents make the old incentive problem more common because an agent will pursue the target with no social friction and no instinct to stop and ask whether the target is the point.
The fixes he points at are human ones. He cites Airbnb tracking bookings and host cancellations together, and HubSpot clawing back sales commission when a customer churns early. Both are paired metrics, where the cheap way to win on the headline number costs you on the second number. He says the same playbook is what you will want for your agents, though the post does not work through what that looks like in an agent harness.
Why it lands here
The shape matters for anyone running Jev as a cheap scorer inside a loop. A confidence gate, an auto-close rule, a router that picks the small model when a score clears a threshold, these are all numbers an agent or a pipeline can be optimized against, directly or by whoever is tuning the thresholds. If the only thing you log is how many items cleared the gate, you have built the bounty and not the measurement.
For people building on it, treat this as a prompt to check what your agent's success metric can be satisfied without doing the work, not as evidence that it already has been. Isenberg is describing a failure mode, not reporting one. The useful move is to go look at whether yours has a second number attached, the cancellation to the booking, before you decide the framing does not apply to you.
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.
