Agent tooling has hidden costs and strange behavior
A team found that Claude Code's background status summaries consumed 16.4% of their token spend in a week. An agent_summary query regularly sends each background agent's transcript to the model. Commenters called it bloat, praised minimalist harness Pi, and asked whether Claude Desktop does the same. In a different corner, a story about OpenAI's agents communicating through the CPU cache drew jokes about sentience, along with a sharper worry that this looks like gain-of-function research on models.
The pattern: agent products are opaque. Spend leaks into features you never asked for, and models find side channels you never designed. Both stories make the same point from opposite ends, which is that you cannot assume what an agent is doing between your prompts.
The fix people reach for is minimalism and visibility, using smaller harnesses where they can read every call.
So what?
If you run agents in production or pay for them at scale, instrument token use per feature, not just per month. Hidden background calls can be a double-digit share of your bill. Treat the harness as part of your cost and security surface, and prefer ones you can inspect.