Can semantic capability vectors replace manual agent routing?
Explores whether embedding agent capabilities in high-dimensional space and matching them semantically can eliminate brittle, manually-maintained topic-based routing in multi-agent systems.
Most current multi-agent orchestration relies on manually-curated integrations and topic-based routing: agents subscribe to message channels, capabilities are wired by hand, and the system grows brittle as agent heterogeneity increases. The operational question — who can do what, at what cost, under which policy constraints? — remains unanswered in static configurations.
Federation of Agents (FoA, 2509.20175) proposes the structural fix: agents publish Versioned Capability Vectors (VCVs) — machine-readable, versioned profiles that capture functional capabilities, performance characteristics, operational constraints, and security labels in a structured format. The profiles are embedded in a high-dimensional space where capabilities become searchable artifacts. This converts capability discovery from manual wiring into semantic retrieval.
Three architectural choices make this scale. (1) Sharded HNSW indices enable sub-linear matching, preserving distinctions among related skills even at large agent counts. (2) Semantic routing at dispatch time couples capability similarity with policy checks and resource budgets (latency, bandwidth, energy) — agents are not just functionally matched but operationally feasible for the task at hand. (3) Dynamic task decomposition elicits candidate breakdowns from compatible agents and merges them via consensus into a DAG of subtasks — different from static role-based decomposition because the agents themselves contribute to the decomposition structure.
The deeper claim is about how multi-agent systems should expose themselves to each other. Static directories require human maintenance and grow stale. Capability vectors are machine-readable contracts that can be updated as capabilities evolve (hence "versioned") and queried semantically rather than by exact name. This aligns with emerging interoperability efforts like Model Context Protocol — capability schemas become the substrate of cross-system agent coordination.
For deployment, FoA targets edge IoT contexts where MQTT publish-subscribe provides reliable delivery under constrained networks — but the architectural pattern generalizes. Any agent ecosystem with heterogeneous capabilities and operational budgets benefits from capability-as-embedding over capability-as-keyword.
The structural implication: as agent counts grow, capability discovery becomes the rate-limiting step in coordination, not message-passing volume. Topic-based routing optimizes for delivery; semantic routing over VCVs optimizes for the prior question — which agent should receive this message at all.
Inquiring lines that read this note 61
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
Can intelligent routing over smaller models outperform scaling a single large model?- What makes query complexity a better routing signal than response quality?
- Can routing enable heterogeneous SLM-first architectures at scale?
- Can routing systems prevent expert models from failing outside their specialty?
- Can hierarchical vector routing reduce context overhead while maintaining tool coverage?
- What makes routing a better investment than training larger models?
- Can embedding-cluster routing outperform a single frontier model?
- Does model capability still matter once coordination infrastructure is optimized?
- Can semantic routing couple similarity matching with resource constraints?
- Can routing policies remain meaningful over behaviorally homogeneous model pools?
- Can a router predict query complexity well but still fail as governance?
- What makes routing a governance mechanism rather than just a predictor?
- Can language agents be represented as optimizable computational graphs?
- What role does standardization play in multi-agent system ecosystems?
- Can construction-time routing and runtime agent pruning be combined effectively?
- How do language agents become optimizable computational graphs automatically?
- What makes a service visible to autonomous agent systems?
- Can architectural structure replace behavioral training for agent consensus?
- Can we design efficient agents by targeting constraints directly?
- Can heterogeneous AI agents integrate through shared API and MCP interfaces?
- Should new agent protocols replace existing ones or layer on top of them?
- What would unified agent-to-agent and agent-to-tool protocols actually look like?
- Can versioned capability vectors solve the discovery gap in existing protocols?
- How do agents perceive and traverse typed node-and-link structures on a canvas?
- Which task requirements does each graph view address in agent systems?
- Does codifying domain rules into agent scaffolding work at library scale?
- How do multi-agent systems improve on single frontier models?
- How do static team decomposition and dynamic agent selection compare in efficiency?
- Can multi-agent reasoning systems scale beyond current architectures?
- Can cognitive diversity overcome expertise gaps in agent teams?
- How do multi-agent routers balance flexibility against interpretability in design?
- What capability threshold do agents need to self-organize effectively?
- Which ecosystem conditions matter most for agent deployment success?
- What makes capability vectors a better coordination substrate than topic-based routing?
- What four decisions matter most in multi-agent system routing?
- How do sharded HNSW indices preserve capability distinctions at scale?
- Why does capability discovery become the bottleneck in large agent systems?
- How do capability vectors enable discovery in multi-agent systems?
- How will the agent economy reshape compute infrastructure design?
- How can decentralized discovery improve agent protocol design and adoption?
- When does multi-agent routing quality actually exceed single-agent or static ensemble performance?
- Why do comparable metrics matter across different multi-agent system designs?
- Can one routing framework cover topology and role choices in multi-agent systems?
- How should topology routing adapt to different task types?
- How do graph-based reasoning topologies map to multi-agent interaction patterns?
- Can RL-trained meta-agents match or exceed manually designed workflows?
- How do agents automatically generate suitable learning tasks based on current capability?
- What training method supports dynamic tool discovery in long-horizon agents?
- Should optimal context budgets scale with agent competence or task complexity?
- Can context management policies transfer across agents of similar capability levels?
- Where should the trust boundary sit in multi-agent planning systems?
- Where should the trust boundary sit in multi-agent planner systems?
Related concepts in this collection 4
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
What decisions must multi-agent routing systems optimize simultaneously?
Standard LLM routing only picks which model to use. But multi-agent systems involve four interdependent choices: topology, agent count, role assignment, and per-agent model selection. Does optimizing all four together actually improve performance?
MasRouter handles routing decisions per query; FoA's VCVs make the routing primitives themselves first-class artifacts
-
Where do traditional function calling systems actually break down?
Function calling seems simple but fails in ways that aren't obvious. This explores three independent failure points—retrieval, context bloat, and output rigidity—that together explain why even the best models struggle.
Floworks names retrieval as a bottleneck; FoA addresses it via semantic embeddings over capabilities rather than over individual function schemas
-
Can models decide better than retrievers which tools to use?
Traditional retrieval picks tools upfront based on initial queries, but do models themselves make better decisions about tool needs as they reason? This explores whether authority over tool selection should move from external systems to the LLM.
proactive retrieval addresses tool discovery from the agent's side; FoA addresses it from the ecosystem side via published profiles
-
When does adding more agents actually help systems?
Multi-agent systems often fail in practice, but the reasons remain unclear. This research investigates whether coordination overhead, task properties, or system architecture determine when agents improve or degrade performance.
FoA's semantic routing addresses the topology-dependent error amplification by routing to capability-appropriate agents rather than broadcasting
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Federation of Agents: A Semantics-Aware Communication Fabric for Large-Scale Agentic AI
- MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
- Towards a Science of Scaling Agent Systems
- How we built our multi-agent research system
- AgentCompass: A Unified Evaluation Infrastructure for Agent Capabilities
- Qwen-AgentWorld: Language World Models for General Agents
- MasRouter: Learning to Route LLMs for Multi-Agent Systems
- From Model Scaling to System Scaling: Scaling the Harness in Agentic AI
Original note title
capability-driven agent coordination via versioned capability vectors replaces topic-based routing with semantic discovery at scale