To figure out how to break down a task for an LLM agent, look at the architecture in your governance data. You are building **Liminality**, a question-decomposition and routing MCP server. The "best way" isn't a generic theory. It must fit your infrastructure and goals.
Here is the decomposition strategy for your system:
### 1. Use Existing "Flows" as Templates You have established flows that show successful decomposition. Map new tasks to these patterns: * **Conceptual/Explanatory Tasks:** Use the pattern from *"Explain how DNS resolution works"* and *"Explain how DNS resolution works and what a recursive resolver does."* This implies **hierarchical decomposition**: split the topic into core concepts, then sub-concepts, grounding each in factual retrieval. * **Operational/Workflow Tasks:** Use the pattern from *"How do I migrate my Postgres database... and notify my team by email?"* This implies **sequential, tool-bound decomposition**: 1. Identify the primary action (migrate DB). 2. Identify the secondary action (notify team). 3. Map each to specific connected toolkits (e.g., `postgres` tools, `gmail`).
### 2. Constrain Decomposition by Available Assets Your logic must fit what you actually have. The "best" decomposition is **executable** with your current assets: * **Acme CRM / Hubspot:** If the task involves customer data, decompose into CRM-specific queries. * **Airtable / Asana:** If the task involves project management, decompose into record creation or status updates. * **Google Suite (Drive, Calendar, Slides, etc.):** If the task involves content or scheduling, decompose into specific API calls for these services. * **physea.ai / mcp.physea.ai:** Use these for external data grounding or specialized processing required by sub-questions.
### 3. Address the "Owner-Only" Bottleneck (Strategic Decomposition) You noted that *"93% of our MCP traffic is owner-only (acquisition is the bottleneck)."* Your decomposition strategy should prioritize **clarity and reusability** to reduce friction for users (both human and AI) adopting Liminality. * **Best Practice:** Decompose tasks into **reusable routes** (as mentioned in your goal: "returns reusable routes"). Break a complex task into sub-questions that can be cached, shared, or called by other agents, rather than one-off, ephemeral queries.
### 4. Critical Missing Information To provide a more precise decomposition algorithm for your specific use case, I need to know: * **What is the specific task you are trying to decompose right now?** (e.g., Is it a customer support query, a data analysis request, or a content generation task?) * **What is the target audience for this decomposition?** (Are you optimizing for human readability or for AI agent parsing efficiency?)
### Pending Action Before I can provide a concrete decomposition plan for a specific task, please confirm: 1. **Which specific task** would you like me to decompose using the Liminality framework? 2. **Do you want me to generate the sub-questions and map them to your existing toolkits** (e.g., Airtable, Gmail) as a pending action for your review?
*(Note: I will not execute any tool calls or create new flows until you confirm the specific task and approve the decomposition plan.)*
This answer was worked out by Liminality — Physea's engine that decomposes a request, grounds each part to a real tool, and returns a reusable, checkable route.
Connect it over MCP: https://mcp.physea.ai/mcp · physea.ai