Agentic LLM stacks refer to architectural patterns for building autonomous AI agents that integrate language models as reasoning engines within larger systems. Model selection within these stacks has evolved toward pragmatic, cost-aware approaches that evaluate frontier and open-source models based on total cost of ownership (TCO) rather than raw capability metrics alone. This shift reflects maturation in the AI infrastructure landscape, where deployment efficiency, operational costs, and performance characteristics have become primary decision drivers.
An agentic LLM stack combines multiple components: the language model itself, retrieval systems, external tools and APIs, memory mechanisms, and orchestration logic 1). The model selection layer determines which LLM serves as the core reasoning component. Historically, this decision defaulted to the most capable frontier models from providers like OpenAI or Anthropic. However, contemporary practice increasingly evaluates open-source alternatives—such as Kimi K2.6 and similar models—based on whether their performance-cost ratio justifies the operational trade-offs compared to commercial offerings 2).
The architecture typically includes a planning and reasoning layer that handles task decomposition, an action execution layer that interfaces with external tools, and a feedback loop that incorporates results back into decision-making 3). Model selection directly impacts latency, throughput, and cost at each stage.
Modern model selection frameworks evaluate multiple dimensions beyond base model capability. Inference costs include per-token pricing and latency overhead; operational costs encompass infrastructure, maintenance, and integration complexity; opportunity costs reflect performance differences that may require additional inference rounds or context length expansion 4).
Open-source models like Kimi K2.6 have become viable defaults in certain agentic scenarios because they offer favorable cost-to-performance ratios when deployed on optimized infrastructure. A system requiring 100,000 daily inferences may see substantially lower total costs with a smaller open-source model despite slightly higher per-query latency or context window limitations. This economic analysis represents a maturation beyond earlier periods when frontier model capability was the dominant consideration 5).
Despite improvements in frontier model efficiency—where providers have reduced inference costs and improved speed—the relative cost advantage of open-source alternatives has not diminished proportionally. This creates a competitive environment where model selection depends on specific deployment characteristics: batch processing scenarios, real-time applications, multi-turn conversational agents, or specialized domains with domain-specific model fine-tuning requirements.
The shift toward open-source as viable defaults reflects infrastructure maturation—better serving, quantization techniques, and optimization frameworks have made self-hosting economically feasible—and capability convergence—where open-source models have reached performance thresholds sufficient for many agentic applications without requiring the latest frontier capabilities. This represents a move away from monolithic “one model for all tasks” approaches toward heterogeneous stacks that mix models based on task-specific economics 6).
When implementing agentic stacks, engineers must evaluate: context window requirements for multi-turn agent interactions and tool invocation history; tokenization efficiency affecting practical throughput; fine-tuning feasibility for domain adaptation; API availability and latency characteristics for agent responsiveness; and regulatory constraints on data residency and model governance.
Open-source models provide advantages in customization, data privacy, and cost predictability. Frontier models offer advantages in reasoning capability, instruction-following consistency, and multi-modal integration. Optimal agentic stacks often employ mixed strategies: using smaller open-source models for straightforward tasks and frontier models for complex reasoning steps that cannot be decomposed into agent subtasks 7).
The pragmatic approach to model selection reflects broader infrastructure trends: economics-driven architecture, where cost optimization drives design decisions; open-source commoditization, where freely available models with reasonable performance become baseline assumptions; and specialized model ecosystems, where different models serve optimized purposes within larger agent systems.
As frontier model costs stabilize and open-source model quality continues improving, agentic stack design increasingly focuses on model routing—dynamically selecting between available models based on query complexity—and hybrid deployment—operating heterogeneous model ensembles that optimize total system economics. This represents maturation of AI operations engineering as a field parallel to software infrastructure management 8).