Alibaba Cloud's Higress team open-sources HiRoute to route agent tasks by model

What HiRoute does

HiRoute is described as a local engine for intelligent agent routing and collaboration. The idea is that users keep the agent they already know, let economy models take care of routine steps, and bring in a primary model only for demanding stages. When a task suits a separate worker, it can be delegated to another agent.

Alibaba Cloud says the problem comes from fragmentation. Models differ in strengths, agents differ in workflow (the post names Codex and Claude Code as tools people favor), and compute is spread across subscriptions, API keys from different providers and local services. A long task may begin with research, move to design and implementation, and end with validation and difficult repairs, so a single model is rarely the best fit for every step.

The team previewed the open-source plans in a keynote at AGNTCon + MCPCon China and formally released the project at the Apsara Conference. According to the post, configuration, routing and observation run locally, while the connected model services perform inference.

Two layers: task routing and model routing

The post separates two layers that can be switched on independently or together. In its words, "Task routing answers 'who should do this?' Model routing answers 'what capabilities does this stage need?'"

For model routing, users can combine subscriptions with lower-cost APIs. If a supported subscription such as Codex is available, HiRoute can connect it through CLIProxyAPI (CPA) and place both sources in one routing plan. The example given puts Qwen 3.8 Flash in the economy group and a stronger model in the primary group.

HiRoute also supports decision models, which judge the category and difficulty of the current task and can assess the previous execution stage. Users can connect a model from Bailian (Alibaba Cloud Model Studio) or TypeSafe's Jev through OpenRouter, without deploying their own decision service. HiRoute then applies the user's thresholds, model groups and candidate order. Simple tasks go to the economy group, while complex tasks, or work that follows a stage below the user's standard, go to the primary group. The criteria and assessment prompts can be edited, and the approach is not limited to coding: the post describes separate writing and review branches for an article, each with regular and primary models.

Cache continuity and failover

Agents work through repeated tool calls, and context keeps growing. The post notes that high KV Cache hit rates reduce input costs and response latency. HiRoute's context continuity feature, implemented by ContextHold, recognizes calls whose history continues within the same user turn, keeps the earlier routing decision and prefers the current eligible model so the cache can keep being reused.

If context compaction or history rebuilding breaks the old prefix, HiRoute reassesses the model at that point and continues along the new prefix. A new user request triggers a fresh decision.

Other features described in the post:

• When a candidate is rate limited or unavailable, HiRoute tries the next eligible candidate within the plan's boundaries, and stops explicitly when candidates run out.
• The Sessions and Model performance views show which model ran each stage, why it was selected, how much it used and whether the stage was assessed.
• Users can draw on their own execution records, not only general benchmarks, when building the next routing plan.

A cost test, and what it means

Alibaba Cloud presents a first case under the heading "over 90% lower cost". The task was a structured engineering research job: investigate six messaging systems, produce 60 source-verification cards for each, and deliver a decision memo on Kafka and SQL consistency. The company says a large volume of research supported a few critical decisions. It ran the task in three modes, including all GPT-6 Astra and HiRoute mixed routing. The source text available here ends before the full comparison, so detailed results are not covered.

For developers and teams building their own agents, the release offers an open, locally run way to mix subscriptions, APIs and agents they already use. Because the code is open, the post says, developers can review routing behavior, connect their own workflows and improve the system through real tasks. The cost claims come from Alibaba Cloud's own test and have not been independently verified.