What changed
Agent Swarm 2.11 adds persistent per-agent Docker workspaces, retrying key-limited tasks on another key in the same profile, webhook filters, completion events over MCP and GitHub App tokens for worker git access. It also fixes parallel-workflow successor loss and adds request-composition manifests.
Why it matters
Persistent state, credential failover and workflow resumption define what survives an agent crash and which identity resumes work. The new controls improve durability and traceability, but stale workspace data, duplicate successors and token scope become release-gate concerns.
What to do
Upgrade a disposable stack, persist one agent workspace, fail a key and interrupt a parallel workflow. Verify one successor path, bounded retry, workspace ownership, request manifest, webhook filtering, GitHub token scope, terminal retention and rollback.
Caveat
The signed stable release is maintainer evidence, not an independent durability or security assessment. Persistent volumes and retention can preserve sensitive state, while profile keys, GitHub Apps, webhooks and MCP events require least-privilege configuration.