There's a number on my token usage dashboard I didn't think much about until it started moving too fast. Once I was using very heavy context and my sessions got up past roughly 500,000 tokens, usage was climbing at almost double the pace.
There have to be a lot of people frustrated with their token usage on the big frontier-model subscriptions. This post is about the one thing that made the biggest difference for me.
What was eating the tokens
When I looked closer, the pattern was plain. Agents keep re-reading the same thing, over and over. The longer the session runs, the more of that you're paying for.
You need to eliminate that.
The move
Here's the one I'll give away. When a session hits around 200,000 tokens, it ends, and a new session starts with a proper handover prompt. It happens automatically.
Automating that stretched my sessions out significantly. It's been a game changer.
It's not the only thing I do to cut usage. There are a bunch of other tricks. This was a big one, and it's the one worth talking about. The rest I'm keeping.
The handover is where it goes right or wrong
A fresh session is only useful if it knows what already happened.
Get the handover wrong and you get hallucinations. You get the agent re-reviewing things that were already done, already discovered, already fixed. All sorts of things.
Which means a bad handover brings back the exact waste you started the new session to get rid of, and adds mistakes on top. The handover prompt has to be done properly, or the new session isn't a fresh start. It's a confused one.
When not to switch
Some moments you don't cut a session off. In the middle of a PR merge, or anything of that caliber, you want it to play out.
The automation has that built in. It doesn't switch sessions in the middle of something like that.
How I got here
Trial and error, and research. Tracking everything I was doing, and monitoring the agents with proper governance. That tracking tells you a lot. It's what shows you how to navigate this instead of guessing.
You're always looking to improve. When you're studying new technology, that's the job: study, learn, and listen more than you act. Then take what you've learned, A/B test it quickly, and set up the infrastructure properly.
So if you're frustrated with your usage and want to know what to do tomorrow, it's one word.
Observe.