Principle AI Product Building Architecture Business Models

The context flywheel is a Day 90 moat — Day 0 comparisons are misleading

Point-in-time capability benchmarks miss the compounding advantage: on Day 0 a raw model matches your product, but by Day 90 accumulated context creates an unbridgeable gap

@izzymiller (Izzy Miller, Hex) — Building AI Agents for Data Analytics · Apr 10, 2026 · 11 connections

“On day zero, Claude with a Snowflake connector is probably just as good. But by day 90, you’re operating in a totally different ballgame.” Hex’s competitive moat is not point-in-time capability — it’s the compounding flywheel: user work → artifacts accumulate → admin observability surfaces patterns → guides and semantic models improve → agent gets better → more user work. This flywheel gets stronger over time and is hard to replicate because it requires both time and active usage.

This operationalizes Context is the product, not the model with a time dimension: context isn’t just a product differentiator, it’s a compounding differentiator. Day 0 demos are misleading for both builders and buyers. This connects to Accumulated agent traces produce emergent world models — discovered, not designed — the emergent world model IS the Day 90 advantage. The implication for evaluation is profound: Evals are behavioral pressure vectors, not neutral measurements — poorly chosen evals distort agent development needs a time axis. Izzy is building “Metric City” — a 90-day simulation to test this compounding: Day 0 accuracy ~4%, Day 90 with Sonnet 4.6: 24% (target: 100%), the benchmark described in Long-horizon evals test compounding behavior, not point-in-time accuracy. And for competitors: Proprietary feedback loops create moats that widen with every interaction — each day of usage widens the gap, making late entry increasingly expensive. FurtherAI names the Day 90 endpoint in insurance: The intelligence lives in the workflow, not the model — and a model can't simply read it — over thousands of production runs the workflow stops being a script and becomes the carrier’s “operating memory,” which is exactly the compounding asset this Day 90 framing describes. Nadella frames the same dynamic as firm strategy: You can offload a task, or even a job, but you can never offload your learning — the Day 90 gap is accumulated learning you can’t outsource — and names that loop as the firm’s hill-climbing IP.

Connected Insights

References (7)

→ Accumulated agent traces produce emergent world models — discovered, not designed → The intelligence lives in the workflow, not the model — and a model can't simply read it → Context is the product, not the model → Evals are behavioral pressure vectors, not neutral measurements — poorly chosen evals distort agent development → Long-horizon evals test compounding behavior, not point-in-time accuracy → You can offload a task, or even a job, but you can never offload your learning → Proprietary feedback loops create moats that widen with every interaction

Referenced by (4)

← You can offload a task, or even a job, but you can never offload your learning ← The data flywheel is a UX problem — only vertical workflow surfaces can capture the knowledge ← The intelligence lives in the workflow, not the model — and a model can't simply read it ← Accumulated agent traces produce emergent world models — discovered, not designed