This video walks through how Aisle Tasks can cut token consumption by up to 50x compared to a fully agentic setup, by trading runtime reasoning for a deterministic Python script that only calls the model where a decision actually needs to be made.
The 50x is not a tuning trick, it is a difference in shape. An agent re-reasons about the whole problem on every turn, and every tool call carries that reasoning overhead with it. A task defines the process once, in code, and spends tokens only on the steps that genuinely need a model's judgment - so the bill tracks the actual work instead of the model's exploration of it.
For more on why this distinction matters as models get more expensive and more variable, see Token Burn is a Design Problem and Announcing Tasks.
Video transcript
To start out with, I've created a project called Asana Performance and I've connected an Asana connector. So, from here I can ask fetch the count of tasks completed by each assignee in the AI work board. That's the board we use to keep track of our tickets that are in progress. The chat is going to go ahead and execute the Asana tools looking at workspaces, looking at projects. It has found the board.
It's looking at the tasks and it will fetch and summarize that data for me. Jumping ahead, it's found the data, it's rendered a table, it's rendered a nice bar chart. Um and if we go back, we can see the work it did to sort of find all that information and collate it. So, if we come up to the top of our chat, we have a little drop-down that says 66,000 tokens were used in this investigation. Now, I know that the system prompt for this project is about 15,000 tokens, which means we spent about 50,000 tokens doing this calculation.
I think we can improve on that. So, what I'm going to say is uh please write me a short ticket I can copy and paste uh to have a developer implement this report using the Asana API. And what I'm looking for are just some loose requirements that explain how you would go about implementing this. That looks pretty good. So, I'm going to copy this information.
And that's really all we need. And rather give rather than give this to a developer, I'm going to head over to tasks. Create a new task and I'm going going it fetch completed count by assignee. We're going to enter the AI mode and I'm going to say please implement the following. I'm going to ask it to output structured JSON and no chats.
So, sometimes when we create tasks, we have it output a chat, which is a nice way to present a report that a user's actually going to read. In this case, we just want the task to output structured data that the agent is going to receive and interpret. Fast forwarding, we have a script that looks pretty good. Indeed, it's outputting structured data like we asked. I don't see a chat.
I'm going to accept this work and close the AI editor. There's one thing we need to fix, which is to connect our Asana credential. This gives the task permission to access that Asana instance, and then we can create task. So, let's test it out first. I'm going to run it right from here.
And we're going to watch the logs and make sure it's doing what it should. All right, that looks pretty good. This data matches pretty much what we saw in the chat. So, heading back over, let's go to manage our tools again. And we're going to add this new task as a tool.
So, now this project chat can call that task directly. Let's try it out. I'm just going to ask it to fetch the completed count by assignee. Okay, you'll see it's running the task. And that was pretty quick.
Here's our output. It matches what expect to see. And what's really nice is if we come up to this indicator here, we've only used about 17,000 tokens. Like I mentioned before, the system prompt is about 15, 16,000. So, we were able to retrieve this data in less probably less than 1,000 tokens.
Uh it was 50,000 to determine this data before by actually querying Asana directly. So, that is a massive savings. And if this is a kind of calculation or report that we want to do all the time, this is going to save us a lot of money. And it's also going to extend the life of the chat thread. I can use this data to ask other questions.
And as this compounds over multiple tools, you can have very, very productive, very fast sessions that are really token efficient.