Argentic

Quick answers

What is Argentic?

Argentic is a hosted engine that runs the agent loop: it calls the model with your key, keeps the agent's state and memory in a journal, and hands each tool call back to your code to carry out. The prompt is rebuilt from the journal under a cap, so it stays small as the history grows. You set the budget, and every run comes with proof of what ran. Kill your agent mid-action and it resumes without repeating a finished action.

Does the agent write summaries of what it did?

No. When the prompt reaches its budget, the engine rebuilds it in code from the journal and keeps the newest turns word for word. Neither the engine nor your agent writes a summary of the older ones; they stay in the journal and stay searchable.

Do you store my model key?

No. The model key comes with each wake in the X-Model-Key header, is used for that call only, is sent only to the endpoint you named, and is scrubbed from any error text. It is never logged or written to the journal.

What happens when the beta ends?

On December 16 Prototype starts counting calls: 1,000 agent calls a month, with the same limits as the beta. Production is open now if an agent needs more headroom. Owners and admins of every billed organization get an email on November 16, 30 days before the change.

What counts as an agent call?

One message to your agent and the work it does to answer it: its model turns and the tool calls in them. Each call is paid for once, even if it was cut off and resumed. A call that fails before doing any work is free.