Set a limit before the agent has spent anything
The cheapest minute you will spend on this product is the one setting a daily
limit on a new agent. Loop detection catches a runaway relative to the agent’s
own baseline — a new agent has no baseline, which is exactly when a limit does
the work. Budgets and guardrails.
One agent per job
Seats get cheaper as you add them, and an agent’s workspace is its memory of what
it has been doing. Two unrelated jobs in one agent share that memory and every
credential in it.
Fund the agent, not only the account
Account balance is not spending power. Money has to reach the agent’s provider
key before a model call can draw on it, and an agent with a full account behind it
and an empty reserve will still sit silent. The Balance panel shows all three
places at once.
Keep credentials out of the config file
Anything the agent needs at runtime goes in
environment variables, where it is encrypted and
write-only. A token pasted into the config file is readable by anything that can
read the config.
And when you rotate one, revoke it at the service that issued it — removing it
here does not reach into a container that is already running.
Take backups out yourself
The volume survives restarts, redeploys and upgrades. It does not survive
deleting the agent. If the workspace holds something you would miss, copy it out
through the terminal.
Read the state, not the vibe
The product distinguishes between measured zero and nothing measured, and it
does it in words: issued on first deploy is not a zero balance, could not
be read is not a working key, and we could not confirm this payment is
neither success nor failure. Where a screen refuses to give a verdict, the
refusal is the information.Last modified on August 31, 2026
