Qwen Code 0.23.3 can delegate a subagent to Claude Code — with one important boundary
Qwen Code v0.23.3 adds ACP-based external subagent delegation, with Claude Code as its first adapter, and raises qwen serve's default workspace registration limit from 25 to 256.
QwenLM published Qwen Code v0.23.3 on September 10, 2026. The release lets a Qwen Code subagent delegate one turn to an external coding agent through the Agent Client Protocol (ACP); its first wired adapter is Claude Code. It also raises qwen serve’s default workspace-registration limit from 25 to 256. (Source: Qwen Code v0.23.3 release notes, 2026-09-10)
Key facts:
v0.23.3was published on September 10, 2026.- The external-subagent feature uses ACP.
- Claude Code is the first external agent adapter wired into Qwen Code.
qwen servenow accepts 256 registered workspaces by default, up from 25.- The standard daemon defaults to 800 total sessions when its registration capacity is above 25.
What this means if you build with coding agents
This is delegation, not a provider switch for the entire Qwen Code session. The Qwen session remains the parent: its transcript, permissions bridge, artifacts, and virtual subagent session stay there while an individual worker is run by an external ACP process. That is useful when a workflow needs a different coding agent for one bounded task without re-keying the workspace or moving the whole session to another backend. (Source: QwenLM PR #11003, merged 2026-09-10)
The more important practical change is fail-closed behavior. The PR says older builds could ignore an unknown executor field and silently run the work in-process under Qwen instead. In v0.23.3, malformed executor blocks are rejected, and constraints an external process cannot honor are refused rather than quietly dropped. That protects an intent such as “use Claude for this worker” from becoming an unlabelled Qwen run.
For a multi-repository local-coding setup, qwen serve now has more room before it rejects a workspace registration. Operators can set QWEN_SERVE_MAX_WORKSPACES to an integer from 1 through 256. Do not treat 256 workspaces or 800 sessions as a hardware-sizing recommendation: the release PR explicitly says the runtimes are still created eagerly, and does not present those policy limits as measured safe concurrency or memory capacity. (Source: QwenLM PR #11515, merged 2026-09-10)
If you are pairing Qwen Code with a local endpoint, start with the Qwen3.6 local coding guide for model and llama-server setup, then keep tool execution constrained using the llama-server built-in tools safety guide. The new delegation path does not remove the need to review what an external worker is allowed to run.