The agent's day-to-day toolkit got an overhaul. Exploring is faster, editing lands where you mean it, and commands no longer get a free pass to run in the background.
Faster exploration
Searching across large repos is dramatically faster, and results come back cleaner — no more surprises from quoting or shell rules.
Commands stay bounded
Commands now run with a clear time budget and better platform guidance. They finish or fail more predictably instead of hanging in the background.
Forgiving edits
Edits tolerate line-ending differences and prefer small, uniquely-anchored matches. The agent lands on the line you meant, not the one that looked similar.
Approval before every command
Long-running commands no longer get auto-approved. Every one goes through the same approval flow as a foreground command.