Parallel agents

Read-only sub-agents called in the same turn run concurrently under a shared cap (max_parallel_agents, default 4) with live progress. A sub-agent's report has its paths checked on disk; a missing one is flagged at the top with the real file of that name.
Background and follow-up. A sub-agent called with background: true returns at once; its report is handed to the model when it finishes, and a turn does not end while one is still out. Every finished sub-agent returns an id: calling the tool again with resume asks that same agent a follow-up with everything it already read (the last 6 are kept).
parallel_tasks (guarded mode) fans write-capable workers out into separate git worktrees, merges them in dependency order into zerd/<run>/integration, detects overlap, re-runs real conflicts on top of the merged state, and flags them for review instead of resolving them by guessing. --parallel lets the model plan the split; --graph tasks.json is explicit. It only pays off for genuinely independent work. Measure before relying on it.