Subsumption Pattern Learning (SPL) hierarchical foundation model agent architecture with MCP integration for email classification. Includes three-layer reactive-tactical-deliberative servers plus baselines (LangChain, AutoGen).
This MCP server exhibits significant definition quality gaps. While tool names follow a clear verb_noun convention (validate_email, get_world_state_summary, match_email, learn_pattern_from_l2, list_patterns, classify_email), descriptions are present but minimal (avg 80 chars, below the 194-char baseline). All six tools have schemas with typed parameters and descriptions, but the descriptions are severely under-developed, most lack WHEN to use the tool, WHAT it returns, or domain-specific guidance. Parameter descriptions exist but are generic ('Agent identifier', 'Email sender address') without format constraints, valid ranges, or LLM-facing guidance. Output schemas are undocumented in the visible code, the method signatures show return Dict[str, Any] with no published structure. Error handling is absent: classify_email returns a dict with 'suppressed' and 'reason' fields for error cases, but does not guide the LLM on recovery or next steps. The deliberative_server code shows budget exhaustion and suppression logic, but errors are not categorized as retryable, user-fixable, or fatal. No tool declares permissions, audit trails, or rate limits. The composition is reasonable (six single-purpose tools operating on email data), but the interface is not optimized for chat, agent_id is required on every call, forcing boilerplate parameter passing rather than context-aware defaults.
Classify email with LLM, respecting budget and suppression rules.
Return a minimal world state summary for orchestration.
Validate email inputs and update world state with any violations.
Learn a new pattern from L2 output if confidence is high enough.
List patterns and metadata for debugging.
Match email against healthy patterns, updating counts.
Output schemas undocumented. Tools return Dict[str, Any] with no published field definitions. LLM cannot predict available fields or plan downstream tool chains.
Parameter descriptions are generic and lack LLM-facing constraints. 'Email sender address' does not clarify format (RFC 5322?), validation rules, or examples. 'Confidence score 0-1' is correct, but most params have no range or format spec.
Tool descriptions are under 40 characters (well below 194-char baseline) and lack WHEN-to-use guidance. 'Validate email inputs and update world state' does not explain when to call this vs. classify_email, or what violations are detected.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | F | 47 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 46 | - | v1 |
Error handling is not actionable. Errors return dicts with 'reason' strings but do not guide the LLM: budget_exhausted offers no retry path; suppressed_by_state offers no unsuppression mechanism; mcp_error_deliberative offers no troubleshooting steps.
No permission declarations. Tools modify world state (validate_email updates violations, learn_pattern_from_l2 learns patterns, classify_email updates suppression) but do not declare required scopes (read:email, write:patterns, write:state).
agent_id is a required parameter on all tools. This forces boilerplate; better pattern is server-side injection via initialization or request context. Agents should not manage agent_id in every call.