before you plug a new tool into your agent:
- check what it can do, not what the readme says
- read only unless it needs write
- treat its output like a stranger talking
- spend, post or delete still needs your ok
- know how to switch it off fast
each new tool is a new way in
your agent's api key is a wallet. treat it like one:
- hard spend cap at the provider
- one key per agent, never shared
- alerts on usage spikes, not the bill
- keys in a secret store, not the prompt
- rehearse a rotation before you need it
a leaked key drains like a hot wallet
your agent did something weird. first moves:
- pull its keys before you debug
- pause it, don't delete it. you need the logs
- list every approval it holds
- check what else shares its box
- save what it just read
fix the prompt last. the damage happens during "wait, why"