Welcome to the New Frontier of Agentic AI | Wiz, Merge, You.com & More | RAISE Summit 2026

Model Selection and Routing Strategies

• Organizations are abandoning single-model reliance in favor of multi-model routing architectures similar to multi-cloud strategies.
• Cost sensitivity is shifting from a secondary concern to a primary decision factor as usage scales and rate limits are reached.
• Studies indicate no correlation between token spending volume and organizational productivity, with cost reduction being the only significant efficiency driver.
• Model selection is increasingly driven by governance, regulatory compliance, and data residency requirements rather than raw performance metrics.
• Blackbox advises against per-step model routing for coding tasks, noting that such over-engineering yields negligible performance gains compared to phased approaches (planning vs. execution).
• Blackbox's internal benchmarks suggest a hybrid approach combining a closed-source reasoning model (e.g., Nvidia Omniverse Ultra) with a high-performance open-source coding model (e.g., Qwen 2.5-High) maximizes output.
• Wiz's "Cyber Model Arena" benchmark demonstrates that model performance varies significantly across specific security tasks, necessitating task-specific model selection.

Security, Privacy, and Infrastructure

• End-to-end encryption at the token level is identified as the highest security standard for agentic inference, surpassing dedicated GPU or air-gap deployments.
• There is a strong market preference for American open-source models and deployments; Chinese models are only favored if fully deployed within a US air-gapped environment with end-to-end encryption.
• Ed (Pincap TieDB) emphasizes that engineering harnesses, tooling quality, and sandbox environments are becoming more critical to success than the underlying model choice itself.
• Viz (Wiz) highlights that the "mode" or intelligence resides in the data graph and architecture, which must remain stable regardless of model volatility.
• Blackbox notes that for code-specific workloads, the market effectively consolidates around five top models, reducing the need for complex routing across hundreds of options.
• Vue.com warns that models are often trained on their own agentic frameworks (e.g., Claude + Code), creating portability issues when switching to open-source frameworks like Ollama.

Evaluation, Variance, and Reliability

• Industry benchmarks relying on single numbers are insufficient; quantifying variance is essential due to non-deterministic agent behavior and inconsistent task difficulty.
• Agent inconsistency stems from multi-turn settings and tool usage rather than model capability alone, leading to reliability risks at scale.
• Vue.com's research indicates that benchmarks often measure agent inconsistency (harness issues) rather than true model capability or task difficulty.
• Wiz employs "context-aware tool selection," filtering available skills to match the specific task to reduce context usage and improve tool selection accuracy for smaller models.
• Strategic tooling partnerships (e.g., access to licensed financial data) can enable standard models to outperform frontier models on specific benchmarks like FinComp without model tuning.

Market Trends and Future Outlook

• The industry is transitioning from a scientific problem to an engineering challenge focused on infrastructure stability and orchestration.
• No single "mythos" event has yet occurred that renders current architectures obsolete; the pace of change is continuous rather than punctuated by one-time shifts.
• Open-source models are rapidly closing the performance gap with closed-source frontier models, particularly in defense and planning scenarios.
• The primary barrier to enterprise automation remains the need to build trust, governance, and compliance layers to safely deploy autonomous workflows.
• User behavior is shifting from short chat interactions to long-horizon agentic workflows requiring extensive tool usage and context management.
• Future differentiation will likely stem from unique tool integrations and data partnerships rather than the base model alone.

Strategic Recommendations for Builders

• Prioritize the quality and integration of tools over the selection of the underlying model for most use cases.
• Build infrastructure that abstracts model choices, allowing for seamless switching between providers to mitigate vendor lock-in and leverage the "best of breed" for specific tasks.
• Focus on developing stable data models and graphs that remain constant despite the volatility of underlying model weights.
• Organizations should deploy their own open-source models on dedicated GPUs for privacy and security, utilizing end-to-end encryption.
• Enterprises must build "bridges" involving governance and guardrails to facilitate the adoption of disruptive agentic technologies.
• Developers should increase attention on open-weight models, as they are becoming capable of handling complex tasks that previously required frontier models.
• The current low cost of development encourages a "build first" mentality, allowing for rapid iteration and risk-taking in agentic applications.