How to Burn Tokens Without Burning Out

If you started using AI agents hard in the last year, you may recognise this. You sleep less than you used to. You feel more capable than you have in years. Things you would have called absurd a year ago now look possible. That feeling is real. It also costs something.

I ran a lot of agents at the same time, across several projects. I moved between threads and terminals. I checked results, answered questions, and started the next prompt before the last one had finished. Each result was impressive enough that I wanted the next one straight away.

The results were too good to stop

I was sleeping four or five hours a night, and it did not feel like a problem. The results felt like the reward. I was always thinking about what to ask next. In the day I read to my two toddlers about magic, and then I watched something like magic happen at work, with real upside for Juma Labs. Clients came in. Work got done. I was proud of it, and my endorphins were at their peak the whole time. Everyone seemed happy, or I thought so, and all of it ran alongside my health.

It was like Aladdin and the lamp, with a genie who grants wishes. Claude was the lamp and Codex was the ring. I kept thinking of what to ask for next, and they kept granting it.

The notification became a leash

Some weekends I sat with my family at the park and ran Claude sessions from my phone. A notification came from one of three threads. It was easy to answer, so I answered it. Then another one came.

Answering was easy. The cost showed up elsewhere: in my temper, in the questions from my two toddlers that I didn't answer, and in the times I wasn't there when they needed me. Four or five hours of sleep a night has side effects too. My concentration suffered, and my temper got shorter.

Threads started to mix

At some point I was answering the wrong question in the wrong session. Threads sat too close to each other. I merged them, and the merges turned into conflicts. What I had called flow was context switching at a higher speed.

A session runs for 15 to 30 minutes. When you start one, you have what looks like time to do something else. That time seems free, but it overlaps with the other sessions that are still running. The free minutes were never free.

The worst part was how much I trusted the defaults. Agents went into auto mode and took the recommended answer. I accepted it without working out what it would do. So I made a rule: if the agent's questions are too complex for me to understand, it's time for bed.

There is no getting on top of it

Eventually I realized this was no way to live. I still felt in control, or it felt like it. I was not driven by the vibe, or so I told myself.

I would love to tell you that I changed. I did, but not as much as I wanted to. The clearer lesson is that there is no catching up with AI. There is no "just this one prompt, then I'm going home," and no "let me do one more thing while this runs." New results arrive about as fast as you clear the old ones.

Keeping up is the wrong goal. Choosing your pace is the right one, and the practices below helped me do that.

Dizzy from AI, you make slop

When you have been running agents for too long, you stop checking what they produce. The output looks finished, so you call it done. Too much AI at once makes you dizzy, and the work you ship gets worse.

The bigger cost comes when that output leaves your desk. A partial, uncertain result sent to a teammate becomes their problem. They spend their own time working out what is true, and that cost lands on them.

Pace the sessions, not the agents

• Read each result to the end before you start the next thread. Sharing it with the team can wait.
• Write the thinking down first. I keep my notes in markdown until the prompt is ready to go in one shot, which means fewer rounds of follow-up and fewer endless loops.
• Group threads by topic. I use Orca as a terminal orchestrator. Each topic gets a worktree, and results open when I decide to look at them, not when they pop up on my screen.
• Name your terminals so you can tell them apart. My Warp terminal windows had one-word names. Five characters was all the title could hold, and that was not enough to remember what a session was for.
• Collect questions, then answer them in a batch. A batch costs less attention than a stream of single questions.
• Talk to people in the afternoon. Talking with the team gives my head a rest from the agents, and talking about what I go through keeps me sharp. I am not always helping them. Often I am just listening and sharing.
• Let the limits stop you. The hyperscalers have usage limits. I use it up, then I rest until it resets or I switch to another one.
• Keep the night for sleep. Nothing important happens at night that you wouldn't wake up for in the morning. If it's important, wake up early, and don't eat into your pillow time.

The short version

1. Read each result to the end before you start the next thread.
2. Write the prompt in markdown first, then send it in one go.
3. Group threads by topic, and open results when you choose.
4. Batch your questions, and answer a notification only when it blocks the thread you're in.
5. Nothing important happens at night that you wouldn't wake up for in the morning. If it's important, wake up early and keep your pillow time.
6. Talk to people, and respect the usage limits.

Tonight, pick one thread, read its result to the end, and close the laptop before you open the next one. The results will still be there in the morning.

I'm still working on this, and I expect to keep working on it. AI agents are unfolding in front of us, and magic is going to be everywhere. Everyone will be a wizard. We will all have to learn to rest too. It's about the journey, not the destination.

Glossary

Terms

Git worktree

A second working directory attached to the same git repository, so another branch can be checked out at the same time without cloning again. Each worktree has its own files and index.

Tools

Claude

Anthropic's family of large language models and the assistant built on them, available in apps and through an API.

Codex

OpenAI's coding agent, which runs in the terminal and is also offered in IDEs and the cloud.