# Agent performance, cost, and failures in SQL

Source: https://www.hotdata.dev/industries/agent-analytics
Site index: https://www.hotdata.dev/llms.txt

Search, join, and roll up every session, tool call, and model call.

## How data flows

Data comes from:
- Agent sessions
- Tool call logs
- Customer database

Hotdata runs in your VPC (Full-text search, Vector search, Percentiles), and the data is used by:
- Eval dashboards
- Cost reports
- Debugging agents

## Common workloads

- **Session analytics.** See where sessions finish and where users drop off. 5×: Faster than Snowflake, with 10 concurrent clients.
- **Quality and evaluation.** Compare outcomes and latency before and after a prompt or model change. [4×: Faster than BigQuery, at every concurrency level.](/blog/benchmarking-concurrency-on-hotdata-snowflake-and-bigquery#hotdata-and-bigquery)
- **Usage and cost per customer.** Tokens and spend by agent, model, customer, and day.
- **Investigating failures.** Find the sessions that went wrong and the steps that led there.

## How Hotdata fits

- **Append sessions as they finish.** Append sessions and tool calls as they finish. See [Load modes](/docs/push-data#load-modes).
- **Search transcripts in SQL.** Find sessions by the text of a prompt, response, or error, filtered in the same query. See [Full-text search](/docs/sql#full-text-search).
- **Find similar sessions.** Find sessions like the one you are looking at. See [Vector search](/docs/sql#vector-search).
- **Percentiles and time buckets.** Latency percentiles and time buckets in plain SQL. See [Aggregate functions](/docs/sql-functions-aggregate).
- **Join with the rest of the business.** Cost per customer is one query, in PostgreSQL, DuckDB, Snowflake, or ClickHouse SQL. See [HotSQL](/docs/sql).
- **Many readers at once.** With 10 concurrent clients, Hotdata runs 5× faster than Snowflake.

## Get started

The [quick start](/docs/quick-start) gets a table into Hotdata in a few minutes.