SAGE Router bets the hard part of multi-agent work is knowing when not to switch

Count the terms in the objective and the scope becomes concrete. One reward, a predicted probability of success, then penalties for cost, latency, risk, context-transfer loss, coordination overhead and uncertainty, with an exploration bonus on top ^[10]. That is eight quantities to estimate for every candidate arrangement ^[1], and six of the eight price something that has not happened yet: what a swap would cost, not what the incumbent has already spent.

Hard estimation for a component that, by design, never sees the result. The prototype returns a routing decision and does not transmit tasks ^[4], so the context transfer it penalizes is one it does not perform. The repository does split capability into a global trust score and trust conditioned on the specific requirement, then calibrates against outcome evidence ^[13], which is the right correction to reputation that travels across work types it was never earned on. The evidence, though, has to arrive from an execution client this project does not include ^[4]. Until that loop closes, the penalty weights are assertions rather than measurements.

The scheduling side is where the design earns its keep. Budget and deadline are checked at the team level after the team has been constructed, rather than read off an individual agent profile ^[9]. That ordering is correct and expensive: it means the router cannot prune infeasible candidates from agent cards, and has to build, schedule, then reject. Independent requirements spread across different agents can run in parallel, while assignments concentrated on one agent serialize ^[8], so whether a deadline holds depends on the critical path estimate ^[7] and not on how many agents were recruited. Beam search over team prefixes instead of greedy team-building ^[14], with credit for covering requirements the existing team does not cover rather than for adding another strong overlapping profile ^[15], is the piece I would keep if the rest were thrown out.

The dev.to writeup is careful about what this does not demonstrate: it is not evidence that open agent networks have solved reliable delegation ^[16], and no benchmark results accompany it. What it does mark out is the seam. A directory can say which agents exist and cannot say whether changing the active team midway through a constrained task is worth the disruption ^[18]. A handoff gains specialist capability and loses accumulated task context; collaboration covers missing skills and creates coordination overhead ^[6]. A router that optimizes apparent competence alone will keep picking the impressive profile ^[17], which is harmless until the incumbent is the one holding the context that decides the task ^[12].