Comparing notes: compact tool output, structural navigation, and end-to-end agent speed

[rs545837]

Hey RTK team, I'm Rohan, working on Sem and Weave at Ataraxy Labs: structural code navigation and editing/merging for agents.

We've been comparing agent sessions against native search/read/edit workflows. One lesson I got is that fewer tool calls or smaller responses don't automatically mean a faster successful session. Missing context can lead to follow-up reads, and validation/retry loops can eat the savings.

Curious what you've learned:

• Which filters help end-to-end task completion most, beyond reducing output bytes?
• How do you test whether compression makes an agent miss something or need extra recovery calls?
• Have you explored combining RTK with structural navigation tools, particularly for build/test diagnostics?