Scaling

Concern	Approach
Concurrent reviews	Scale workers; raise `STEW_MAX_CONCURRENT` carefully
HTTP / webhooks	Scale API (JWT is stateless)
UI traffic	Scale UI static pods / CDN
Queue depth	Optional broker + KEDA
Multi-tenant isolation	Org-affine worker pools + strict sandbox

Worker notes

Prefer one job at a time per process when applying org runtime env
Use Docker/k8s sandbox in multi-org shared pools (STEW_TENANT_ISOLATION=strict)
Graph MCP is embedded — scale workers, not a separate graph service