Sprix SAGE Router, a library that reroutes AI agent tasks while they run, reached 3,814 GitHub stars in 20 days

Sprix SAGE Router, an open-source Python library that decides whether a running AI agent task should continue where it is, recruit collaborators, or hand off to another agent, has reached 3,814 GitHub stars since its repository was created on 18 August. The counts come from the GitHub REST API, retrieved 7 September: 3,814 stars and 434 forks in 20 days.

SAGE stands for State-Aware Graph Exchange. It sits above the Agent2Agent (A2A) protocol, which gives multi-agent systems a shared format for agent cards, messages, tasks and artifacts. On top of that discovery layer, SAGE scores three routes for a task already in flight: the current agent keeps it, a small team joins while the current agent keeps ownership, or a specialist takes it over completely. The score accounts for completed work, reusable artifacts, observed quality so far, remaining work, failures, budget and deadline, and the decision comes out as one auditable objective.

Its own benchmark, with the caveats printed next to it

The repository's benchmark replays 1,000 task checkpoints over five random seeds and compares seven strategies under the same registry, budget and deadline. In those runs, the progress-aware router scored a utility of 0.298, against 0.085 for always continuing and 0.290 for always handing off, and it missed deadlines on 23.3 percent of tasks against 34.4 percent and 30.5 percent for the two fixed strategies. Those are the project's own numbers from a synthetic replay, and the README says so directly: it labels the project an early-stage research preview, states that the synthetic studies are no substitute for real trace data, and lists what a production rollout would still need, including calibrated evaluators, authenticated agent identities, signed capability metadata and monitoring.

The README describes the project as an open-source research output of Sprix AI, which it presents as the agent-to-agent initiative of the Chinese company 屿智同行. That description is the project's own account of itself. The code is MIT-licensed, requires Python 3.10 or newer, runs a test suite in CI, and was last pushed on 28 August.

Teams building multi-agent systems today mostly decide ownership once, at the start: a planner assigns the task to one agent, and the assignment stands until the task finishes or fails. Work already completed, half-built artifacts and a slipping deadline usually enter the picture only when a person looks in. A library that turns the continue, recruit, or hand-off choice into a single scored decision gives builders a concrete place to put that logic, and an audit trail for why a task moved. The evaluation is self-run and synthetic, and the honest labelling is part of what makes the repository worth a look: the method, the seeds and the caveats ship with the numbers, so a team can rerun the replay before trusting it. 3,814 stars in 20 days for a research preview says the mid-task question is one builders recognise.

Source

Primary source: wang2122/sprix-sage-router on GitHub. Star and fork counts from the GitHub REST API, retrieved 2026-09-07.