November 2026 Summaries

Hex reports reducing the cost per multi-message conversation with its analytics agent by nearly 75%, from $6.44 in January 2026 to $1.64, as usage and token consumption grew substantially. The company attributes the reduction to lower-cost models, including OpenAI’s Luna and Kimi 2.7 for suitable tasks, and to agent-harness improvements that reduced unnecessary token use through better caching, tool selection, error handling, and specialized subagents. Its DataBench benchmark and Context Studio evaluation tools are intended to help customers compare model quality and cost on analytical work and on their own data environments. Hex also introduced a code-first 4.0 rewrite that reduced the agent’s tool set to roughly 40 tools, with projected thread-cost savings of 20–40%, and plans to further lower costs by allowing the agent to retain knowledge across related analyses rather than repeatedly rediscovering information.

During a summer internship at Hex, Deena Sun built Hero, a recursive self-improving agent system that analyzes real user interactions to identify failures, clusters recurring issues into root causes, and helps generate fixes that can improve the same agent platform it relies on. Hero complements offline evaluations by observing live usage, using analyzer agents to inspect individual trajectories and aggregator agents to distinguish meaningful patterns from noise before turning findings into actionable pull requests. The project reflects a broader shift in AI-assisted engineering, where rapidly changing models, prompts, tools, and codebases make software increasingly temporary, while Hex reports sharply rising development throughput. Sun argues that the durable value of research is not necessarily its code artifacts but the questions, experiments, lessons from failures, and conceptual foundations they produce. Although AI makes implementation faster and accelerates software’s ephemerality, the author maintains that human curiosity, lateral thinking, and judgment remain central to deciding what to investigate, making ideas and accumulated learning the lasting “stones” beneath constantly shifting technical “sand.”

Business intelligence tools are increasingly incorporating generative AI to let leaders investigate KPI changes through plain-language follow-up questions rather than relying solely on dashboards or data-team requests. The comparison evaluates Hex, ThoughtSpot, Power BI, Tableau, Sigma, Looker, and Qlik primarily on user autonomy, governance, answer provenance, saving and sharing capabilities, pricing, and integration with existing technology stacks. Hex is presented as a shared AI-native workspace combining conversational investigation, reusable applications, and governance tools, while ThoughtSpot emphasizes search-driven exploration, Power BI and Tableau align with Microsoft and Salesforce environments, Sigma focuses on editable warehouse-native workbooks, Looker relies on governed LookML models, and Qlik combines agent-based analytics with capacity-based pricing and access controls. The guide notes that AI capabilities and limits vary widely, including query caps, separate infrastructure costs, permissions requirements, model-development effort, and beta or preview features. It recommends piloting finalists using a real production KPI, testing a progression from basic measurement to driver analysis, verifying security and data definitions, inspecting how answers were generated, and determining whether results can be saved and reused before making a purchasing decision.