Minimal coding-agent tool surfaces

Simon Willison's llm-coding-agent shows a compact coding-agent design built around line-numbered reads, exact-string edits with diffs, bounded shell commands, file globs, regex search, and scoped approvals. Why it matters: The useful signal is restraint: smaller tool surfaces are easier to audit, test, and recover when an agent makes a bad edit or command choice.

Try this: Before adding more autonomy to a coding assistant, define the smallest read, search, edit, and command tools it needs, then require visible diffs and test output for every change.