Qwen Code Weekly: Agent Teams, Durable Scheduling, /cd, and More

Qwen Code released versions v0.18.0 through v0.18.3 this week (June 18, 2026), merging over 100 PRs. The updates address several common workflow friction points: parallel multi-agent collaboration, scheduled tasks that survive restarts, and the ability to switch working directories without restarting the CLI. This article breaks down the key changes, how they work, practical use cases, and trade-offs.

What Changed: Four Versions in One Week

The week’s releases span four versions (v0.18.0–v0.18.3) and cover five major areas: Agent Team parallel collaboration, durable /loop scheduling, /cd directory switching, a collapsible thinking display, and an MCP approval gate. Smaller improvements include faster context compression (/compress-fast), permission bubbling for background agents, and workflow enhancements for parallel subagent execution.

Agent Team: Persistent Parallel Collaboration

How It Works

Previously, /fork could spawn background agents, but each agent worked in isolation with no communication channel. You had to manually combine results from multiple agents. The new Agent Team mode (experimental, opt-in) changes this. Once enabled via experimental.agentTeam setting or QWEN_CODE_ENABLE_AGENT_TEAM=1 environment variable, the model can create named teams and spawn multiple “teammates” that work in parallel. Teammates and the leader exchange messages, share task lists, and the leader consolidates a final unified report.

Key design difference: teammates are persistent. They are not one-shot subagents that disappear after completing a task. After finishing, they wait for the next assignment. The model gains new tools: create/delete teams, message teammates, and create/update/list shared tasks. The Agent tool adds a name parameter to spawn long-lived named teammates.

Practical Use Cases

• Have three reviewers review code in parallel—one focusing on error handling, one on race conditions, one on edge cases—and get a consolidated report.
• Coordinate a small team to build a bookmark application, with teammates automatically reporting to the leader upon completion.

Limitations

Agent Team is experimental and disabled by default. Teammates are model-driven; the quality of collaboration depends on the underlying model’s ability to use the team tools effectively. There is no manual control over teammate behavior beyond the initial prompt. The feature is opt-in and may not suit all workflows.

Durable /loop: Schedule Tasks That Survive Restarts

How Durable Scheduling Works

Previously, all /loop scheduled tasks were session-bound—they vanished silently after a restart. Durable /loop persists tasks to disk, allowing them to survive restarts. When a request implies persistence (e.g., “check my PR every hour”), the task is saved per project to ~/.qwen/tmp/<project-hash>/scheduled_tasks.json. This file is outside the workspace and won’t appear in git status. On next launch, tasks are automatically recovered.

Default behavior remains session-only: temporary reminders like /loop 5m remind me do not write to disk. Only requests that clearly need persistence create durable tasks.

Edge cases are handled:

• One-time reminders that expired during downtime: on restart, the model confirms before executing.
• Missed periodic tasks run one catch-up cycle after startup, then resume normal rhythm.
• When multiple sessions exist for the same project, only one triggers persistent tasks. If the holder crashes, another takes over within seconds.
• Periodic tasks auto-expire 7 days after the last trigger.

Durable Scheduling Use Cases

• “Check my PR every hour” continues running after restart—no need to reschedule.
• Missed periodic tasks automatically catch up once on restart.
• Temporary reminders like “remind me in 5 minutes” still perish with the session.

Durable Scheduling Limitations

Only tasks that imply persistence are saved. Session-only reminders remain ephemeral. The 7-day auto-expiry may be too short for some use cases, but prevents indefinite accumulation. Multi-session coordination is automatic but may have race conditions in rare edge cases.

/cd: Switch Directory Without Restarting CLI

How /cd Works

The new /cd <path> slash command lets you change the current session’s working directory without exiting the CLI. It validates the target directory, handles trust confirmation, updates workspace roots and cwd service, migrates current session files (transcript, runtime, worktree) to the new workspace session directory, and refreshes system/directory context. Path completion works as expected.

If the target path is untrusted, the trust confirmation flow triggers, consistent with first-startup trust behavior. The command supports navigating up and down within a session; it does not automatically detect directory changes.

/cd Use Cases

• /cd ../another-project switches to an adjacent project; session files auto-migrate.
• /cd /path/to/parent-repo navigates up and down within a session—no restart needed.

/cd Limitations

Only explicit /cd commands trigger directory switching; there is no automatic detection. The feature is part of the CLI and not available in headless or daemon modes. When moving between projects, verify the active workspace before running commands.

Other Notable Improvements

• Thinking Collapsible: Model reasoning blocks no longer dominate the screen. During streaming, a 4-line fixed-height scrolling window shows live reasoning with a timer. After completion, it collapses to one line showing total duration (e.g., “Thought for 15s”). Use Ctrl+O to expand and view the full reasoning. The thinkingDisplayMode setting was removed in favor of this single collapsible mode.
• MCP Approval Gate: Project-level MCP servers (declared in .mcp.json or workspace settings) no longer auto-connect. They must be approved before connecting. Approval decisions are stored in <QWEN_HOME>/mcpApprovals.json, indexed by project root and server name, and bound to a config hash. Changing a server’s command, args, URL, environment, or headers changes the hash, reverting it to pending. This prevents malicious MCP configs from auto-connecting without review.
• Background Agent Permission Bubbling: Background agents requiring permission confirmation for tool calls (e.g., git push, rm) can now bubble the request to the parent session’s background tasks UI. Set approvalMode: bubble in subagent config. After approval, the agent continues; after denial, it exits normally. This avoids the previous behavior where background agents auto-denied, forcing re-runs in the foreground. Note: bubble is subagent-only; in headless/ACP/SDK contexts it still auto-denies.
• /compress-fast: A pure rule-driven context compression command that does not require an LLM call. It forces microcompaction: cleans old tool results and media content (keeping the most recent N), strips all thinking parts from model replies. Zero API calls, zero token overhead, instant response. Useful for quickly reclaiming space in long sessions.
• Workflow Enhancements: Workflows now support parallel() and pipeline() for concurrent fan-out (opt-in via QWEN_CODE_ENABLE_WORKFLOWS=1). Workflow P3 adds schema, agentType, model selection, and worktree isolation for subagents.
• Declarative Agent Frontmatter: .qwen/agents/*.md files now support permissionMode, maxTurns, color, mcpServers, and hooks for per-agent configuration.
• Tool Output Truncation: Per-tool budgets (shell 30k, grep 20k, agent 32k) prevent oversized output from blowing up context. Excess is saved to temp files.
• Other Fixes: OOM protection, SSH clipboard, Windows printf command missing, cursor movement at wrap boundaries, and more (see full changelog in PR references).

Limitations and Trade-offs

While the new features address real pain points, they come with trade-offs:

• Agent Team is experimental and model-dependent. Not all models may effectively use the team tools.
• Durable /loop adds a new system file in ~/.qwen/tmp. Tasks are project-specific and do not survive project hash changes (e.g., if the project directory is renamed).
• /cd changes the active workspace inside the CLI. It is useful for navigating between project directories, but it is not available in headless or daemon modes.
• MCP approval gate adds a step when opening new projects, but it’s a necessary security measure.
• Background agent permission bubbling works only in interactive CLI mode; headless and SDK contexts still auto-deny.
• /compress-fast sacrifices semantic quality for speed; use /compress when summary quality matters.

Concrete Takeaway

If you are a hands-on Qwen Code user, start with Durable /loop and /cd—they have the most immediate impact on daily workflow. Enable Agent Team if you frequently run parallel tasks that benefit from inter-agent communication. Review the MCP approval gate to understand the new security posture. For long sessions with many tool calls, try /compress-fast to reclaim tokens instantly.

These updates show the Qwen Code team systematically tackling workflow interruptions: scheduled tasks should not vanish on restart, directory switching should not require a CLI restart, and background agents should not waste a run due to missing permissions. The changes are practical, well-considered, and worth adopting.

Sources

• Original source

AI-assisted summary compiled from the sources above, reviewed by a human before publishing.