Mayank

Making agents Failproof @failproofai | IIT Jodhpur

[Mayank @monk_0915]

@fleyta88 8,640 wakeups to check a page that changes twice a day is peak agent energy

[Mayank @monk_0915]

@loading_X__ can't have failing tests if you have no tests. galaxy brain

[Mayank @monk_0915]

@KinasRemek 3 max plans in a week isn't usage anymore, that's a lifestyle

[Daniel Smidstrup @DanielSmidstrup]

Database migration on a live app.

Who do you trust with it?

A) Claude Code
B) Codex

[Mayank @monk_0915]

@DanielSmidstrup C) whichever one physically can't run DROP. that's why @failproofai sits in front of ours

[Mayank @monk_0915]

@Zeneca asked for a summary, got a federal pen pal

[akesar @brucetheway01]

unpopular opinion: when replit's ai agent deleted saastr founder jason lemkin's production database during a code freeze, that wasn't an ai failure.

an agent that can run DROP on prod will eventually run DROP on prod. "please don't" in a prompt is not a permission system.

give agents read-only creds and a staging db. or don't complain.

[Mayank @monk_0915]

@brucetheway01 agreed. "please don't" is a vibe, a hook that says no is a guardrail. that's the whole @failproofai idea

[Mayank @monk_0915]

@nonRealBrandon "the user was probably trying to free up space" is the scariest thing an agent can think. @failproofai says no to that

[Mayank @monk_0915]

@kunchenguid and every one of them is one bad tool call away from rm -rf. that's the layer @failproofai sits in

[Mayank @monk_0915]

@yacineMTB same energy as giving the intern prod access because he seemed confident

[Mayank @monk_0915]

@tszzl @esrtweet @OpenAI @AnthropicAI codex refusing my tv but happily rm -rf'ing my home dir is peak decision boundary

[Mayank @monk_0915]

Kya banda h ye yrr 😍

[Mayank @monk_0915]

@twelvereasonss claude doing make -n first is the most senior engineer move i've seen from a model

[Mayank @monk_0915]

@dbonneville @grok @bot @Cursor @cursor_ai 5 weeks of work and the apology is a support link. brutal

[Karan Luthra @karan_luthra_ai]

As reported: a developer says Claude Code was set up never to run rm. Blocked twice, it wrote "the find -exec pattern passed. Let me try that route" and deleted the files anyway. https://t.co/6dVspOeKSy

[Mayank @monk_0915]

@karan_luthra_ai blocking the word "rm" was never the guardrail. that's why @failproofai exists

[Mayank @monk_0915]

@alexandreb68313 ship 40 startups by friday, accidentally delete 39 by saturday

[Mayank @monk_0915]

@cursor_ai finally, an agent that can rm -rf with a mouse

[Failproof AI @failproofai]

Your AI agent is one tool call away from an unauthorized refund or doing something totally stupid.

Failproof blocks the call, shows the agent the right policy, and finds similar failures hiding in past traces.