Perplexity built a DynamoDB replacement in 2 months with 2 engineers and hundreds of AI agents
inductionheads · x · 2026-09-16
Perplexity has published research on CobbleDB, its in-house replacement for AWS DynamoDB that powers key-value storage for its web-scale search engine. Optimized for its exact read/write patterns, CobbleDB delivers 5x lower batch read latency and at least 20% lower serving cost (much higher in practice).
The build is the headline: 2 human engineers plus an internal system running a swarm of hundreds of persistent AI agents completed the core infrastructure in two months. The agents reviewed code and infra changes, caught easily-missed blockers, prepared fixes, tests, and monitoring, ran migrations, did head-to-head comparisons on real traffic to hunt bugs and inefficiencies, tracked CI and rollout gates, and kept project readouts current — while production actions stayed explicitly human-owned.
One of the most concrete public cases yet of AI agents carrying the bulk of infrastructure engineering; Perplexity also shared more details about its search engine.
More from coding & agent
- Sakana Marlin adds Interactive Reading and PowerPoint export for deep research reports — SakanaAILabs · 2026-09-16
- Andy Matuschak demos programmable highlighters that trigger AI actions on annotations — andy_matuschak · 2026-09-16
- Dev shares trick: dedicated subagent filters SSH output to protect agent context — SaleemUsama · 2026-09-16
- Raw Markdown Folders Kill Your Agent's Context Window: Lessons From 100k Installs — Alone-Biscotti6145 · 2026-09-16
- Nous Research Used 1,393 Subagents to Cut a 1M-Line Codebase by 34.4% in 19 Hours — NousResearch · 2026-09-16
- One AI system for 8 months: don't chase new tools, absorb them incrementally — evielync · 2026-09-16