Perplexity built its search database with two engineers and AI agents

Perplexity replaced the database behind its search index and says the core work took two human engineers, hundreds of AI agents and two months.

The system and the numbers

The company published its research on CobbleDB on September 15, 2026. It is a key-value store purpose-built for one workload: the repeated batch reads that serve prepared web page records to Perplexity search. It replaced DynamoDB in production.

The reported before-and-after is specific. Median batch-read latency fell from 31.4 to 5.60 milliseconds, and the p99 from 123 to 24.2 milliseconds. Perplexity’s internal cost model estimates at least 20 percent savings relative to DynamoDB. The company frames the before-and-after as observational production measurements rather than the same requests replayed side by side, and says it also ran a synthetic benchmark against both systems with comparable batch sizes and payload ranges.

Perplexity says it plans to open-source CobbleDB for other teams building AI search, which would let outside engineers check the numbers against their own workloads.

The ratio is the claim

Two months for a production database is the part worth sitting with. This is not a model writing a script. It is core infrastructure with real latency budgets and correctness requirements, and the humans in the loop moved from writing the code to directing a fleet that wrote it.

Whether that ratio transfers is the open question. Perplexity has compute, a narrow and well-understood workload, and engineers who already knew what the system had to do, which is close to the best case for agent-written infrastructure. The open-source release is what will show whether the result holds up outside the conditions that produced it.