Key Takeaways

  • AI agents do not follow human persona models; they fire dozens of exploratory, unpredictable SQL queries simultaneously across multiple databases.
  • Speed determines viability for agentic systems: low latency is mandatory because serial query lag compounds instantly inside automated reasoning loops.
  • ClickHouse scaled past $350 million in ARR by solving high-throughput ingestion bottlenecks, including workloads like Tesla processing one billion events per second.
  • Future infrastructure will cater to software agents that possess their own cryptographic identities, operational permissions, and compute budgets.

The Death of the User Persona

Traditional enterprise software was built for predictable human schedules. A finance manager opened a dashboard at 9:00 AM on Monday, pulled a quarterly revenue report, and closed the tab. Software architects mapped compute requirements directly to these tidy user personas.

Aaron Katz, Co-Founder and CEO of ClickHouse, points out that agentic systems throw this entire design pattern in the garbage. “Software applications were designed for a specific persona that had a role within an organization and their query patterns were very predictable and whether or not you ran a report or you looked at a dashboard, agents don't have personas.”

When an autonomous agent attempts to solve a problem, it does not wait for a weekly sync. It queries five databases at once, evaluates the results, notices an anomaly, and runs forty more queries across three separate schemas. Katz notes: “What's the number one requirement for agent query patterns? Low latency because they're executing dozens of SQL queries simultaneously across all these different systems and the most important requirement is the unpredictability of those query patterns, the responsiveness and the fact that they're much more exploratory than a traditional human report or query.”

Surviving a Billion Events Per Second

When software operates without human bottlenecks, data volume explodes. Systems no longer wait for user clicks; automated agents and connected sensors constantly stream state changes and raw telemetry.

Katz points to extreme enterprise deployments to demonstrate the scale modern infrastructure must handle. “Tesla, for example, is ingesting a billion events per second into ClickHouse. That throughput is unprecedented. And so you need the ability to both ingest that efficiently, store it efficiently, and then be able to query that efficiently at a fraction of the cost of traditional database technologies.”

If an agent requires twenty sequential round-trips to complete an automated action, a 500-millisecond database response creates a ten-second delay. That delay kills the user experience. You cannot build real-time agent workflows on databases designed for batch exports.

Software Selling to Software

The long-term shift is not just how data gets queried. It is who buys the software. Today, human engineers evaluate databases, run benchmarks, negotiate enterprise contracts, and provision clusters. Katz argues this workflow is temporary.

“We can talk about the fact that agents will need to have an identity that they don't have today. They'll need to have a budget.” Once agents have verified identity tokens and spending caps, they will run their own performance tests and select their own tools.

Katz is building ClickHouse around that exact shift. “I'm thinking about a future where the agents actually select the infrastructure stack behind the application and positioning ClickHouse to be the default database for the next generation of applications that agents are building, not humans.”

What to Do With This

Audit your application database logs this week to isolate agent-driven query loops from human dashboard loads. Check the tail latency on multi-table joins when an agent executes parallel tool calls. If your p99 latency exceeds 200 milliseconds per query, your agent will crawl under multi-step reasoning tasks; migrate high-frequency analytical tables to a columnar store built for real-time reads before adding more agent capabilities.