Routing logic that decides when agents should fly solo or tag out

What it does

Sprix SAGE Router is a decision layer that sits above the Agent2Agent (A2A) protocol. It ingests a live task DAG, current execution state, and agent bids, then ranks three routes—SELF, COLLABORATE, and HANDOFF—against hard constraints like permissions, budget, and deadline. The reference implementation assigns DAG roles, estimates critical-path latency, and refines its model after observing actual cost, latency, and per-requirement success.

The interesting bit

Most routing picks the best individual agent; SAGE uses beam search to assemble complementary teams and learns per-requirement trust rather than a single reputation score. It also treats quoted cost and latency as claims to be calibrated, not facts.

Key highlights

• Tri-mode routing evaluates solo, team, and handoff options inside one auditable utility function
• Progress-aware replanning weighs completed DAG nodes, failures, and context-transfer loss before switching strategies
• Requirement-conditioned trust prevents a coding specialist from inheriting credit for research tasks
• Synthetic benchmark shows online learning improving utility over static baselines, though the authors explicitly warn these are not real-world results
• Zero runtime dependencies and a deliberately lightweight policy layer

Caveats

• Explicitly labeled an early-stage research preview, not a production system
• Benchmark numbers are synthetic; the README states they are not evidence of real-world superiority and lists several missing evaluation steps
• The current prototype returns routing decisions but does not transmit or execute tasks over A2A

Verdict

Useful if you are designing agent orchestration and need a constraint-aware, learning router with inspectable rationale. Not what you want if you are looking for a drop-in A2A client or validated marketplace performance.