On big repos, context limits and slow searches are the main drags. Scoping work to the relevant directories and using targeted search instead of broad reads keeps the agent fast and accurate.
A clear project memory file (conventions, build commands, where things live) saves the agent from re-deriving context every session.
Allowlisting safe, repeated tool calls removes permission-prompt stalls — the single biggest speedup on long autonomous runs.
Try it: ClaudHQ on speeding up claude code.