wang2122/sprix-sage-router

What it solves

SAGE (State-Aware Graph Exchange) addresses the problem of mid-execution task rerouting in open Agent-to-Agent (A2A) networks. While discovery services can find agents, SAGE determines who should continue working on a task after it has already started, accounting for work already completed, the quality of partial results, and the cost of transferring context to a new agent.

How it works

SAGE acts as a decision layer that evaluates three primary routing modes based on a utility function that balances success probability, cost, latency, and coordination overhead:

• SELF: The current agent continues the task.
• COLLABORATE: The current agent recruits complementary agents to form a team.
• HANDOFF: The task is transferred entirely to a more specialized peer.

It uses a checkpoint-aware model to estimate how much work a new candidate can reuse from existing artifacts (artifact portability) versus how much they must redo. It also tracks "requirement-conditioned trust," meaning it evaluates agent reliability per specific skill rather than using a single global reputation score.

Who it’s for

Developers building multi-agent systems or A2A marketplaces who need a mechanism to dynamically reconfigure agent teams during a task's lifecycle to optimize for quality and budget constraints.

Highlights

• Checkpoint-Awareness: Estimates remaining work and artifact reuse to prevent unnecessary rework during handoffs.
• Requirement-Conditioned Trust: Tracks reliability per specific requirement/skill to improve matching accuracy.
• Three-Mode Routing: Jointly compares continuing solo, collaborating, or handing off in a single objective function.
• A2A Protocol Integration: Designed to work with the Agent2Agent protocol to map Agent Cards and bids into execution plans.
• Auditable Traces: Provides rationale and alternatives for routing decisions to ensure transparency.