Agent-agnostic token optimization: Save up to 90% with compression, routing and filtering

Agent Native

9 min read

·

Oct 16, 2026

--

3

Listen

Share

Most developer tools are designed for humans staring at a terminal.

They print progress bars, successful tests, repeated file paths, package metadata, ANSI formatting, verbose stack traces, and pages of output that are useful for debugging but rarely useful for an agent’s next decision.

When a coding agent runs those same tools, much of that output gets pushed directly into the model’s context.

Run enough commands and the waste compounds.

• Test suite prints 2,000 lines when three failures matter.
• git diff repeats structural metadata around every change.
• Linter reports the same rule across dozens of files.
• Package manager streams progress the model will never use.

The agent still has to read all of it.

This makes command output one of the most practical places to optimize an agentic system.

Instead of letting raw terminal output flow directly into the conversation, we can:

• Intercept it first
• Parse it according to the command that produced it.
• Remove noise
• Group repeated information.
• Preserve failures and relevant state
• Then return a much smaller representation to the agent