A research team unveiled Agent-as-a-Router, an open‑source approach that treats model selection as a self‑learning agent rather than a static classifier. Their concrete implementation, ACRouter, uses a Context‑Action‑Feedback loop to remember which models succeeded or failed on similar prompts, then routes new requests accordingly. This dynamic memory lets the router adapt to shifting workloads and model updates, avoiding the “information deficit” that plagues heuristic rules and static trained policies.
Static routers only inspect input text and lack any feedback from execution results, leading to three failures: frozen knowledge, poor out‑of‑distribution generalization, and rapid obsolescence when newer models appear. ACRouter’s three‑part architecture—Orchestrator, Verifier, and Memory—stores past outcomes in a vector store, selects the most suitable model for each task, and verifies success by running the output in real environments such as code interpreters or database engines.
In the CodeRouterBench evaluation of roughly 10,000 coding tasks, ACRouter achieved the lowest cumulative regret and cost $13.21 versus $34.02 for a baseline that always used Claude Opus, a 2.6× saving. The framework shines on verifiable tasks like coding or data retrieval, but is overkill for trivial or subjective workloads. The code and model weights are released on GitHub and Hugging Face under Apache 2.0.
