Pattern: three separate projects this week attack agent context cost at three different layers, all with published measurements

GitHub·medium signalXBlueskyLinkedInCopy link

rtk compresses command output at the shell (60-90%), Girder replaces whole-file reads with graph-addressed source (97.85% fewer bytes on its own repo), and declick replaces MCP tool schemas with lazily-loaded CLI verbs (4.1x on 258 tools across nine servers). The shared premise is that the agent's context, not the model, is the binding constraint, and all three publish a reproducible benchmark rather than a claim. If you are picking one, they compose: they hit tool listings, file reads and command output respectively.