UI
Nobody ever loved a database. For half a century, data systems were caught in an endless crossfire between two camps that spoke entirely different languages: humans who craved clean charts to glance at during morning standups, and machines that demanded rigid relational tables they could join. So, we compromised. We built massive warehouses for the silicon, a multi-billion-dollar dashboard industry for the flesh-and-blood, and spent our entire careers trapped in the translation layer between them. Normalization served the machine. Visualization served the human. And the engineer served both, forever.
The Translation Layer Was the Job
Getting a fact landed exactly once, in order, with its history completely intact across systems that constantly fail halfway through—that was always the hard engineering work, and it still is.
Over the decades, those tools earned legendary names: you capture changes right where they happen (Debezium), shuttle them across an event bus that refuses to drop a single packet (Kafka), and make sure you never mistake the pipeline for the ultimate truth. Meanwhile, the human end of the pipe developed its own rigorous science. Data visualization isn't art; it's cognitive psychology. Humans decode position along a common scale infinitely better than angles or pies, which is why the humble bar chart consistently beats the flashy dashboard jewel (Cleveland via Textbook of Usability). As Vega-Lite notes, a chart that requires a separate legend just to be understood has already failed its primary user (Vega-Lite).
The Readers Just Merged
The rise of AI agents changes everything because the audience on both ends of the pipe has fundamentally shifted.
On the machine end, things get radically smoother. An LLM agent will happily maintain contracts, data lineage, idempotency keys, and complex backfill procedures—the tedious correctness work human engineers historically postponed until production blew up. But something even stranger is happening: the two ends are merging.
Agents don't just move your data anymore; they read it. They summarize metrics, judge health anomalies, and trigger actions based on raw output. That means your data representation has to serve a hybrid reader at once—one canonical shape per concept, flattened and denormalized for rapid token digestion, with explicit units built right into field names and zero ambiguous null values that mean three different things.
At the same time, the human end doesn't disappear; it evolves into a verdict layer. The dashboard is no longer a sandbox for endless data exploration. It’s simply the sign-off—the single view a human glances at to confirm the agent-driven system behaved properly. Build those views for immediate human judgment, and keep them ruthlessly honest (WCAG on non-text content).
The old paradigm—warehouses for machines, separate dashboards for humans—assumed our consumers were entirely different species. Now, the entity querying the pipeline is often the exact same agent that built it. Model once, land it cleanly, and show the operational truth plainly.
If you want pipelines without the mess of untyped schemas and broken lineage, I’ve compiled 18 of the bank’s 116 skills into plug-and-play instructions — landing, contracts, backfills, CDC, quality gates, streams, upserts, lineage. You get all 115, no tiers; your agent reads only what matches, because progressive disclosure keeps every task pack small.
Sources: Debezium — change data capture · Kafka — documentation · Textbook of Usability — William Cleveland · Vega-Lite · WCAG 2.2 — non-text content