I had a random discussion in my Telegram group about AI. I noticed that some of them had really high token usage and very long sessions. After some more discussion, I realised that they had fairly complex setups, with lots of coordination between their agents.
Like any other engineer on a weekend, I opened my laptop and fired up two separate Pi sessions to run the same task. One used no subagents and the other used subagents.
Calculate my total token usage across all Pi sessions for the last 7 days,
broken down by session and model, including input, output, cached,
reasoning, and overall totals.
| Setup | Context usage | Time |
|---|---|---|
| No subagents | 63k / 272k (23.2%) | 3m 50s |
| With subagents | 129k / 272k (47.5%) | 8m 8s |
With subagents, context usage increased by 104.7% from 63k to 129k tokens. The elapsed time increased by 112.2% from 3m 50s to 8m 8s.
2x the usage and 2x the time to produce the same result.
Sometimes, it’s not the harness or the models. It’s just our workflows. Not really something new. We went through the phase of bloated MCP. Now we have bloated skills. It makes sense to have bloated workflows too.
It feels unfair to call it bloated too. Some people are more productive with Sublime Text as their code editor. Others are more productive with a JetBrains IDE. To each their own, I guess.
Don’t get me wrong. Subagents are not necessarily bad. I like using them for code review. Asking the author to review their own code, AI or a human, comes with unavoidable bias after all.