sprix-sage-router

### Architecture Review: Sprix SAGE Router Sprix SAGE Router (`sprix-sage-router`) is an open-source research preview developed by Sprix AI at 屿智同行. Operating on Python 3.10+, it targets a specific gap in multi-agent orchestration: dynamic, mid-execution rerouting across open Agent-to-Agent (A2A) networks. Standard A2A protocol infrastructure handles static discovery, Agent Cards, messaging, tasks, artifacts, and transport. SAGE—short for State-Aware Graph Exchange—acts as an analytical decision layer sitting directly above these A2A primitives. Instead of merely asking which agents exist before execution starts, SAGE evaluates task state during runtime to determine who should continue the work. #### The Three Routing Modes SAGE evaluates three distinct routing options inside a single, auditable objective function: * SELF: The incumbent agent retains full task ownership. SAGE triggers this when the current agent’s capability and accumulated execution context are sufficient to finish the job. * COLLABORATE: The incumbent agent retains ownership but recruits a small team of peer agents to fill specific missing requirements. * HANDOFF: Full ownership of the task transfers to a peer agent. SAGE selects this route when the specialist advantage of a new agent outweighs the context-transfer penalty. These three choices are evaluated against a shared, progress-masked action space governed by system permissions, budget caps, and strict deadlines. #### Under the Hood: Pipeline & Continuation Value Rather than treating running tasks as monolithic units, SAGE computes a concrete continuation value. It accounts for work already completed to estimate how much effort a candidate agent would actually need to redo. To calculate this, the router evaluates: * Completed DAG (Directed Acyclic Graph) nodes vs. remaining work. * Reusable artifacts and their portability across agents. * Observed partial execution quality. * Incumbent ownership, past failures, budget consumption, and deadline pressure. System execution follows a multi-step pipeline: 1. Candidate Filtering: Filters agents registered in the network based on hard constraints. 2. Joint Mode Search: Compares SELF, COLLABORATE, and HANDOFF modes across candidate assignments and schedule variations. 3. Plan Ranking: Ranks feasible execution plans under the unified objective. 4. Evidence Loop: Ingests runtime execution evidence to update future routing decisions. #### System Requirements & Integration SAGE is released under the MIT License and requires Python 3.10+. It does not replace transport protocols; instead, it relies on A2A to manage messages, transport, and artifact handoffs while SAGE handles the algorithmic assignment, scheduling, and state-aware decision-making.