Agents use the same AI credits as the rest of Crono — see How the credit system works for enrichment and AI usage for the general rules. This article covers what is specific to Agents.
How much a run costs
The cost of an execution is not fixed: it depends on how much work the Agent had to do. You cannot know it before the run, but you can see it afterwards, per execution, in the Agent's Activity log — which makes it easy to work out your own average after a handful of runs.
What drives the cost:
How many steps the Agent takes. More actions means more credits.
Whether it goes to the web. Runs answered from the data already in Crono are the cheapest; searching the internet and reading websites might increase the credits an execution consumes.
How precise your instructions are. Instructions with no stop condition are the most common reason for an unexpectedly expensive run — see How to write effective instructions for your Agent.
Set a limit per Agent
In the Agent editor, the AI credit limit box on the right sets the maximum number of credits that Agent may consume per month. The box also shows the Refresh cycle — the date your allowance renews.
Leave it empty and the Agent has no limit of its own; it is bounded only by your workspace's allowance.
The limit cannot exceed your workspace's monthly allowance.
Set one on every automatic Agent. It is the safety net that keeps one badly-worded instruction from consuming a month of credits.
What happens when credits run out
When an Agent reaches its own limit, it stops starting new executions. Its card in the Agents list shows a red banner — “You've run out of credits. Increase your limit or wait for your credits to refresh.” — with the refresh date.
The same happens for every Agent when the whole workspace runs out.
Executions already running are not interrupted.
You cannot chat with an Agent that has no credits left.
To get going again, raise the Agent's limit, or wait for the refresh date.
Keeping the cost down
Always end the instructions with a clear stop condition.
Say when the Agent may go to the web, and where to stop looking.
Do not ask the Agent to check whether a list exists — create it yourself and name it.
Start with a narrow trigger, and widen it once you know what a run costs.
One Agent, one job.