1,200 analytics requests a day: How we built Vercel's data agent
summary
D0 is a production data agent built on the open-source framework eve that handles 30,000 questions/month with 45% non-human calls from apps and other agents. It queries Snowflake, runs Python stacks (Pandas, SciPy, DuckDB) in a Vercel sandbox with durable sessions and long-running work (tools can run ~30 minutes), is Git-backed with native evals and observability, and after adopting eve becomes <1/10 the code, 30% cheaper, 40% faster and materially more reliable.
AI-written summary · about
chapters (14)
- 0:00Introduction to D0
- 1:33What is D0 and its features
- 4:20D0 Classic to Diet D0 and Eve
- 5:44Learnings for building agents
- 6:05Production agents need structure
- 7:47Context layer in D0
- 8:53Durable runtimes for agents
- 10:39Security in agent dev
- 12:02Composability in agents
- 13:23Agents calling apps
- 15:28Agents need real SDLC
- 16:51Self-improving agent SDLC
- 17:52Key needs for agents
- 18:13Conclusion and Eve resources
related talks
18m1,200 analytics requests a day: Clay and Vercel's Bet on AI-Native Analytics
56mOpening Keynote
18mAutomating 90% of support tickets: How we built Vercel's support agent
50mOpening Keynote
01mRecap
25mBuilding high performance, AI-native engineering teams
47mThe Agent Era: How AI Agents Are Reshaping Data Platforms
34mRethinking the future of digital transactions