Qwen Code Agent Team multi-provider routing: persistent parallel agents that switch providers at runtime

Alibaba's Qwen Code coding agent shipped v0.18.0 through v0.18.5 between June 18 and June 21, 2026, delivering a set of features that collectively change how teams evaluate open-source coding-agent alternatives and the provider routing decisions that go with them. The headline addition is Agent Team mode — persistent, named agents that work in parallel, message each other, and share task lists — but the release also includes durable scheduled tasks and the ability to switch working directories without restarting. For AI engineering teams managing multi-provider routing, the more interesting detail is structural: Qwen Code already supports OpenAI, Anthropic, Gemini, and Qwen APIs natively, and the v0.18.5 patch explicitly improves provider detection by hostname, making it harder to accidentally misconfigure a non-default endpoint.

What changed in Qwen Code v0.18

The four versions in the v0.18 train introduced changes across three capability areas:

Agent Team mode (experimental). Previously, /fork could spawn background agents, but each agent ran in isolation — there was no communication channel between them, and users had to manually collect and merge results. Agent Team mode lets the model create named teams, spawn persistent teammates, have them exchange messages, and share a live task list. The leader consolidates a final report. Critically, teammates are not one-shot subagents: they can finish one task and wait for another. Enable via QWEN_CODE_ENABLE_AGENT_TEAM=1 or the experimental.agentTeam setting.

Durable /loop scheduled tasks. Scheduled tasks previously disappeared when the session closed. Starting in v0.18, tasks with implied persistence (e.g., "check my PR every hour") are saved to ~/.qwen/tmp/<project-hash>/scheduled_tasks.json and recovered on restart. Tasks that were missed during downtime run one catch-up cycle, then resume. Periodic tasks expire after seven days.

In-session directory switching (/cd). Previously, switching project directories required restarting the CLI. /cd <path> now migrates session context to the new workspace in place, refreshing system and directory context automatically.

Provider hardening in v0.18.5. Among the 30+ fixes in the patch release, fix(core): detect providers by hostname and fix(core): match provider base URL slash variants address edge cases where non-default base URLs were silently misidentified as the wrong provider class. Teams using Qwen Code against custom or proxied endpoints — including AI gateway URLs — will benefit from fewer silent routing mismatches.

Why it matters for AI engineering teams

Qwen Code has been growing as a multi-protocol alternative to single-provider coding agents. The feature comparison published alongside the GitHub repo shows parity with Claude Code on subagents, auto-memory, MCP, plan mode, computer use, and IDE plugins (VS Code, JetBrains, Zed), while adding open-source model and framework weights, runtime provider switching, and IM bot channels (Telegram, DingTalk, WeChat, Feishu).

Agent Team mode is the first step toward persistent multi-agent workflows where task decomposition and parallel execution are handled at the agent level rather than requiring the operator to orchestrate from outside. For teams that already run agent-intensive workloads — long refactors, parallel review passes, continuous background monitoring — this is a routing-adjacent decision: do you wire subagents through a gateway that unifies model spend and context budget, or do you let the coding agent orchestrate locally?

The durable /loop feature has a specific relevance for teams using coding agents in CI/CD or background-monitoring roles. Previously, restarts broke scheduled task continuity. Durable loop removes that fragility, making Qwen Code a more viable option for persistent agentic pipelines that must survive process restarts.

The router/operator angle

The provider-agnostic architecture in Qwen Code creates a specific operator decision point. Qwen Code supports four protocol families at runtime — you can point it at any OpenAI-compatible, Anthropic-compatible, Gemini-compatible, or Qwen-native endpoint without code changes, just configuration. This is structurally similar to how a routing gateway works: the same agent, different providers, different cost and latency profiles.

For teams using an AI gateway as their upstream, the relevant question is whether Qwen Code's provider detection correctly identifies a proxied endpoint as the intended protocol class. The v0.18.5 fix for hostname-based detection and base URL slash normalization reduces the risk that a gateway's URL format causes the client to silently fall back to a default provider class with different behavior.

Two routing implications to track:

• Subagent model routing: Qwen Code's default configuration maps Opus-class requests to V4 Pro and Haiku-class requests to V4 Flash. In Agent Team mode, teammate agents use the CLAUDE_CODE_SUBAGENT_MODEL equivalent. Teams that want the primary agent on a frontier provider and subagents on a cheaper provider — a split-routing pattern — can configure this at the env variable level.
• Durable task budget: persistent /loop tasks continue to consume tokens across sessions. Teams with per-user or per-project token budget policies will need to account for background agent spend that is no longer bounded by session lifetime.

What TheRouter users should watch

Teams currently routing coding-agent traffic through TheRouter who are evaluating Qwen Code as an alternative or complement to Claude Code should watch:

• Whether ANTHROPIC_BASE_URL pointed at a TheRouter endpoint is correctly identified as Anthropic-protocol by the v0.18.5 hostname-detection fix. The base URL variant normalization was a documented fix.
• Agent Team's token consumption pattern — it spawns multiple concurrent agents, each with their own context and model calls. Unified billing across team members' concurrent agent sessions is easier to manage when routing through a single gateway.
• The provider routing docs if you want to evaluate how fallback policies interact with Qwen Code's runtime provider-switching capability. TheRouter's unified billing surface makes it straightforward to compare per-session token spend across provider tiers used by Agent Team teammates.

Qwen Code v0.18 is available via brew install qwen-code or npm install -g @qwen-code/qwen-code@latest. The weekly changelog is published at https://qwenlm.github.io/qwen-code-docs/en/blog/ and the GitHub repository is at https://github.com/QwenLM/qwen-code.