Post

does the orchestrator survive a redeploy mid-run or does state get checkpointed somewhere? asking because i've lost to that exact scenario twice on other platforms

Tried the beta on a scraping pipeline — the built-in retry logic alone replaced about 40 lines of error-handling boilerplate I'd written myself.

The 'your code is the orchestrator' framing is doing a lot of work here. Been stitching together queues and retry logic manually for agents — if Vercel actually abstracts that cleanly, it removes the last reason to touch infra.

That risk inheritance is exactly why we built Runlayer at Vinkius. It acts as a control plane where you can define security policies for MCP servers and skills, ensuring your orchestrator isn't inheriting unvetted risks. Check it out: https://vinkius.com/mcp/runlayer

Specifically the confused deputy pattern: sub-agent can request actions the orchestrator is authorized for, but the sub-task shouldn't need. Fix: scope down at every handoff. Each delegated sub-agent gets a permission subset, never the parent's full scope.