
Claude Agent SDK Framed as Model's Body
The LLM is the brain; the Agent SDK is the body. Tools, files, and the loop around the model make up the harness that lets the model actually do things rather than just talk.

The LLM is the brain; the Agent SDK is the body. Tools, files, and the loop around the model make up the harness that lets the model actually do things rather than just talk.
Having four subagents each attempt to build in their own separate git worktree is described as the wrong approach entirely. The complaint echoes broader frustration that worktree tooling is noisy and outdated.
Large-scale rewrites with Claude start with conversation and hesitation from the model on big deletions, then move to design docs and review docs, before splitting the work into embarrassingly parallel subagent workflows.
Most work doesn't need maximum thinking or effort levels. The team member reports working mostly on medium or light, with a linked writeup explaining what effort actually does under the hood.