I ran a short harness experiment using goose and Reachy Mini.

I built an MCP-based goose extension to control Reachy Mini, then gave it the command: “Look to the right as much as you can.” I measured two things: how long it took for the robot to turn its head to the right, and how many tokens were consumed to complete the task.

Both experiments were run in the same environment, using the same GPT-5.5 High model. The only difference was the number of enabled extensions.

In experiment 1, I enabled a total of 9 extensions, including goose’s built-in extensions and the Reachy Mini extension. In experiment 2, I enabled only the Reachy Mini extension.

The result was quite interesting.

Experiment 1 — 9 Extensions

Tokens consumed: ~14K
Extensions: 9 (goose built-ins + Reachy Mini)

Experiment 2 — 1 Extension

Tokens consumed: ~2K
Extensions: 1 (Reachy Mini only)

The Comparison

• Same command.
• Same environment.
• Same model.

But simply changing the extension setup resulted in roughly 7x more token usage.

Takeaway

The takeaway from this experiment was clear to me: Adding more capabilities to an agent is not always better. For an agent with a clear and specific purpose, exposing only the necessary tools can be much more efficient.

For my goose-based Reachy Mini agent, a single Reachy Mini extension was enough to execute my command well.

This is one of the reasons I like goose. Because goose is MCP-based, I can keep only the capabilities I need and remove everything else. I can even remove basic capabilities like read/write. That made it possible to build the kind of lightweight, purpose-built Reachy Mini harness I wanted.

This experiment made me like goose even more.