Hierarchical planning improves tool coordination across isolated servers
HTN Planning as a Coordination Layer for Multi-Server MCP Tool Orchestration
Artificial Intelligence
Summary
When computers need to work together but are kept separate for safety or organization, it can be hard to direct their teamwork smoothly. The authors introduce a method that designs a clear step-by-step plan beforehand, then runs this plan steadily across different servers without confusing overlap. This approach helps avoid random mistakes and speeds things up by cutting down back-and-forth communication. They tested their system in real science-related tasks, showing it works for complex jobs involving multiple tools and databases.
What this means in practice
- •For bioinformatics platform developers: Control workflows that need real-time, reliable coordination across multiple isolated biological data servers.
- •For laboratory automation engineers: Execute complex robotic experiments spanning multiple isolated control systems with guaranteed stepwise plans and data flow.
Authors
Eliott Jacopin, Éric Jacopin, Koichi Takahashi
Abstract
The Model Context Protocol (MCP) isolates servers by design: only the host can orchestrate cross-server workflows. When the host is a large language model, the resulting orchestrations are non-deterministic, non-reproducible, and pay one inference round-trip per tool call. We present a coordination architecture in which a Hierarchical Task Network (HTN) planner generates a verifiable cross-server plan once, and a runtime middleware executes it deterministically across multiple MCP servers, binding cross-action data dependencies via a template mechanism (\verb|${context.X}|) substituted at execution time. The architecture mirrors MCP's isolation constraint: each compound task decomposes into server-local primitive actions, and inter-server data flow is bound at execution time via JSON-path output extractors. We instantiate the architecture on five HTN domains spanning laboratory robotics, bioinformatics and multiscale modelling, and demonstrate end-to-end execution from a browser-based plan controller against eight live third-party MCP servers querying real biological databases.