Why Analytics Agents Break Differently (hex.tech)

Hex’s Notebook Agent hit a hard limit while analyzing a marketing funnel: compression triggered after ~200k tokens, and the agent lost the actual numbers it needed to reason about conversions. That failure exposed a fundamental difference between coding and analytics agents—code compresses well because abstractions (functions, classes) preserve meaning, but data doesn’t. Tables, JSON blobs and unpredictable cell outputs mean you can’t summarize away values without breaking exploratory reasoning. Attempts to solve this by simply using very large context-window models (1–2M tokens) ran into practical tradeoffs: degraded quality when stuffed with millions of tokens, slower responses, and 2–3x higher cost. Hex’s engineering response reframes context management for data flows rather than chasing bigger windows. They built a cell inventory that reveals notebook structure as a computational graph (cell types, dependencies, outputs), enforced fixed budgets (tool calls limited to 5–10k tokens; notebook context fixed at 15k tokens), and made truncation explicit so agents can request missing data instead of silently failing. These constraints let agents plan, trace lineage, and reason about outputs efficiently, enabling graph-aware navigation and more reliable exploratory analysis. The takeaway for AI/ML practitioners: analytics agents break differently and need context engineering—predictable limits, transparent truncation, and graph-aware tooling—rather than just larger models.