Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Qwen Code 0.27.3: the November 13 build

Written November 13, 2026

Qwen Code 0.27.3 landed November 13. What the release notes list, and where the 0.27 line stands next to the 1.0 preview.

What changed

Qwen Code 0.27.3 was released on November 13, and its notes state there are no known breaking changes. The list underneath is long, and most of it serves the managed and embedded side of the product: the Web Shell, managed agents, and the shared agent loop the team is building underneath both. A terminal-only user will notice a few items; a team running Qwen Code as a managed service will notice more.

Web Shell and sessions

Managed agents and MCP limits

Several items put numbers on managed behaviour. Managed MCP now counts only up to 32 live selections, retains removed entries for recovery, keeps the 16-enabled limit, and caps retained snapshots at 96 entries. New managed-agent SPIs let embedding applications import exact Skill or extension revisions from external sources and store compiled archives in private storage. Managed Sessions store oversized resources and compacted closed journals in object storage, and self-paced /loop wakeups are now stored durably in the session journal, so they survive restarts and takeovers.

Mods gain a display power with a clear boundary: they can rewrite displayed reply rows in the OpenTUI terminal through ui.render, while the stored message text and /copy output stay as the model wrote them. What you see can be reformatted; what the session records cannot.

The loop underneath, and one permission tightening

A run of items builds the unified agent loop without changing current behaviour: internal event-bus primitives, an opt-in AgentLoopOptions.useEventBus, the shared loop enabled on headless runs behind an experimental setting, and AgentCore model rounds and tool batches moving onto the shared loop when experimental.unifiedAgentLoop is enabled. These are groundwork entries. They matter if you embed Qwen Code and track its architecture; they change nothing in a default session.

One fix does change a default session, in the safer direction. In auto-edit mode, tool calls matching permissions.ask rules now require confirmation instead of being automatically approved. If your setup leans on ask rules to keep a human in front of certain tools, 0.27.3 restores what those rules promise. Two environment fixes sit nearby: CLI startup in Docker or Podman sandboxes now creates the sandbox home directory and tolerates a missing $HOME instead of crashing, and manual workspace memory forget now works while automatic memory is disabled.

Where the 0.27 line stands

This is a patch on the stable line, shipped the same week as the separate 1.0 Preview. The 0.27 line is the one the main installation instructions serve, and 0.27.3 keeps its shape: no breaking changes, and the new loop switched on only behind experimental settings. Update for the permission fix and the sandbox startup fix; read the 1.0 Preview as a different track, not as this line's next patch.

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Qwen Code 0.27: agents that ask questions · Qwen Code 1.0 folds in the Qoder harness