Drive several ready-for-agent issues simultaneously. Each runs in its own
disposable, git-isolated sandbox; AGENTIC_CONCURRENCY caps how many run at the
same time. Best when you have a backlog of independent slices and host headroom.
cp examples/parallel-throughput/orchestrator.env .sandcastle/orchestrator.env
# then edit .sandcastle/orchestrator.env → set GH_TOKEN (and AGENTIC_CONCURRENCY)
printf 'CLAUDE_CODE_OAUTH_TOKEN=%s\n' "$(claude setup-token)" > .sandcastle/.env- Each concurrent sandbox is a full container — scale
AGENTIC_CONCURRENCYto host CPU/RAM. Start at 2–4 and watch memory. - On the local tier it's memory-bound: each concurrent sandbox loads the model, so N concurrent ≈ N× the model's RAM. Keep concurrency low there.
- Issues with unmet blockers are skipped; the unblock cascade re-queues dependents as PRs merge, so the pool stays fed without manual relabeling.
afk # autonomous — recommended for throughput
hitl # reviewed — serializes on your approvals, so less parallel in practice