Token
The unit models read and write — roughly three-quarters of an English word. Context limits and API pricing are both denominated in tokens, which makes token discipline cost discipline.
The unit everything is priced in
Models read and write tokens, not words: roughly three-quarters of an English word each, with punctuation, code and unusual proper nouns costing more. Both your context limit and your invoice are denominated in them, which makes token discipline and cost discipline the same activity.
Input and output are not priced alike
Output tokens generally cost several times input tokens, and cached input can cost a fraction of fresh input. A workflow producing long verbose responses can easily cost more than one reading large documents and answering briefly — a distinction invisible if you track only "tokens used".
Where an agent’s tokens actually go
Rarely the user’s question. The bulk is standing overhead re-sent every turn: system prompt, tool schemas, loaded skills, retrieved memory, accumulated history. That is why on-demand skill loading and selective retrieval move the number so much more than shortening prompts does.
In AGNT
AGNT meters usage per run, so a workflow’s token economics are visible instead of a surprise invoice.
Token — common questions
How do I estimate tokens before running something?
Divide the character count by about four for English prose as a rough guide; count code and structured data more generously.
Do tool schemas count against my tokens?
Yes — every declared tool is serialised into the request on each call, which is why large undisciplined toolkits get expensive quietly.
Does AGNT show token usage?
Yes. Usage is metered per run, so a workflow’s economics are visible before they arrive as an invoice.