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: agents that ask questions

Written November 8, 2026 · Edited November 12, 2026

Qwen Code 0.27.0 adds child questions in managed-agent sessions and a task Q&A heatmap. What the release changes for scripted runs.

What shipped

Qwen Code v0.27.0 shipped on November 7 with no listed breaking changes. The headline is for managed-agent sessions: they can now hold multiple child questions at once, answer them in any order, and keep drafts and errors separated per child. A managed agent that used to block on one question at a time can now carry several open questions and work through them as answers arrive.

The same release adds two Web Shell pieces around questions. There is a task Q&A activity heatmap in the composer: clicking a day scrolls the transcript to that day's first Q&A and highlights it, so a long session's question history becomes navigable instead of a scroll to hunt through. And pending Ask User Question panels get an Ask for details action, which opens a side task that explains the options without answering — the explanation is separate from the decision, so reading about an option does not count as choosing it.

What else changed

The GitHub channel gets a Repository Filter in its editor, with allowlist and blocklist modes, and bound pull-request delivery and plain comment delivery become more reliable. The release notes list those as the channel changes; the practical effect is that a GitHub-connected agent is less likely to miss a delivery or to watch a repository it should not be watching.

Underneath, the managed-agent runtime is faster and less prone to deadlocks, with safer session rename, snapshot publication, storage admission, and lock handling. Two security-minded fixes also land: upgrade refusals and Advisor startup errors no longer expose embedded credentials, and empty Harness user info is rejected during start and verify. FileChanged hooks arrive for local non-managed sessions, reporting add, change, and unlink events for watched files and directories, and shell tool edits now handle CRLF and Unicode safely in sed substitutions, with cancelled commands reported as interrupted while keeping partial output.

What it changes for scripted runs

The child-questions change is the one that matters for anyone running managed agents from scripts or a harness. Before 0.27.0, a session with a pending question was a session waiting on one answer. Now a session can hold several pending questions, and the answers can arrive in any order, with each child's drafts and errors kept apart. That makes a managed session easier to drive from outside: a harness can answer the questions it knows and leave the rest open, instead of the whole session stalling on the first unanswered question.

The heatmap matters for the same reason, from the other side. A session that asks questions across days leaves a trail that is hard to review after the fact. Clicking a day in the heatmap jumps to that day's first Q&A, which turns a post-run review — what did the agent ask, what did we answer, where did it guess — into a short check instead of a transcript search.

None of this removes the guessing problem. An agent that can ask several questions can still choose not to ask, and can still answer its own question badly. The release gives the agent a better way to ask; it does not force it to ask. If you run managed agents unattended, the check that matters after upgrading is still the same one: run one session on a real task, watch whether child questions actually fire, and look at what the agent did where it did not ask.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Codex 0.167: an agent command centre · Run parallel subagents in Claude Code · Amp Spaces: orchestrator over parallel threads