Portal Community

When to Use SK Plugin vs Native Octopus

ScenarioRecommendationReason
Standard conversational agentNative Octopus LLM providerLower overhead, simpler debugging, same performance
MCP tool-calling agentNative Octopus LLM providerMCPToolRegistry is purpose-built for this; no SK needed
Multi-step task decompositionSemanticKernelPlugin with PlannerSK planner handles decomposition natively
Existing SK skill library to reuseSemanticKernelPluginAvoid rewriting existing SK prompt templates
Per-step model routing (expensive/cheap)SemanticKernelPluginSK multi-service kernel makes this straightforward
Anthropic Claude or custom LLMNative Octopus LLM providerSK's connector support varies; Octopus ILLMProvider is model-agnostic

Known Limitations

LimitationDetailsWorkaround
SK Planner costEach plan step may require 2–3 LLM calls (plan + execute + verify)Set MaxSteps conservatively; monitor token spend
SK version couplingMicrosoft.SemanticKernel API changes frequently; upgrading may require plugin updatesPin the SK version; test upgrades carefully
No Octopus 4-memory-type integration for SK plannerThe SK planner does not automatically read episodic or procedural memoryInject memory context into the goal string passed to the planner
Streaming supportSK planner does not support streaming responses — the full plan result arrives at onceUse SSE streaming only for non-planner turns
Debugging complexityPlan steps are not visible in the Octopus conversation log by defaultEnable SK's built-in invocation logging and forward to your observability stack

Performance Impact

FeatureAdditional LatencyAdditional Token Cost
SKLLMProvider (no planner)~5 ms overhead vs nativeNone
SK Planner (simple 3-step plan)+1–3 seconds per step~2,000–5,000 tokens per plan
SK Memory bridge< 1 ms per lookup vs direct store accessNone
SK Native functionsSame as MCP tools — negligible wrapper overheadNone
Recommendation. Start with native Octopus features (MCP tools, 4 memory types). Add SemanticKernelPlugin only when you specifically need SK planner, SK skill libraries, or per-step model routing. The added complexity is not justified for standard agent patterns.