What you saw
Why it happened
Manifest caps in-flight requests at 10 per workspace, shared by all its harnesses. Streaming completions hold a slot until the stream ends, so an agent that fires off parallel streams can hit this fast.How to fix it
- Let in-flight requests finish, then retry. Most SDKs handle this automatically with backoff on
429. - Reduce parallelism on the client side. Most agents only need one or two concurrent calls.
- The 10 concurrent-request cap is a fixed guardrail. No env var or setting changes it, self-hosted included. Reduce parallelism as in step 2.