D0
A data science AI agent developed internally at Vercel, used to automate data querying, analysis, and report generation. It serves as a core case study in Vercel's AI agent engineering practice.
Timeline (last 90 days)
Vercel shifted D0 from a multi-agent chained architecture to a single monolithic agent (single AI call, max steps set to 100), allowing it to manage internal state and backtrack to correct errors on its own.
Introducing a file system (bash, read file, write file) free exploration capability for the agent directly doubled D0's evaluation score.
The Vercel team shifted from a multi-agent architecture to a single large agent that possesses full context and can manage its own memory.
A gap exists between high scores on internal evals and real-world performance; the Vercel team achieved 30% on evals but received poor real-user feedback.
After evolution, Vercel's internal data science agent D0 doubled its evals score.
After D0 was opened up to all of Vercel, thousands of queries poured in daily, and the team distilled recent queries into skills via scheduled tasks, accumulating approximately 100 skills to date.
Vercel rebuilt D0 following Claude Code's approach—running in a sandbox, dumping the entire semantic layer into it, and letting the agent freely explore via bash, read file, and write file—which doubled its evals score.
The second version of Vercel's D0 adopts a multi-agent pipeline architecture comprising a query agent, a planning agent, an execution agent, and a reporting agent.
All Facts (8)
Vercel shifted D0 from a multi-agent chained architecture to a single monolithic agent (single AI call, max steps set to 100), allowing it to manage internal state and backtrack to correct errors on its own.
50%UnverifiedIntroducing a file system (bash, read file, write file) free exploration capability for the agent directly doubled D0's evaluation score.
50%UnverifiedThe Vercel team shifted from a multi-agent architecture to a single large agent that possesses full context and can manage its own memory.
50%UnverifiedA gap exists between high scores on internal evals and real-world performance; the Vercel team achieved 30% on evals but received poor real-user feedback.
50%UnverifiedAfter evolution, Vercel's internal data science agent D0 doubled its evals score.
50%UnverifiedAfter D0 was opened up to all of Vercel, thousands of queries poured in daily, and the team distilled recent queries into skills via scheduled tasks, accumulating approximately 100 skills to date.
50%UnverifiedVercel rebuilt D0 following Claude Code's approach—running in a sandbox, dumping the entire semantic layer into it, and letting the agent freely explore via bash, read file, and write file—which doubled its evals score.
50%UnverifiedThe second version of Vercel's D0 adopts a multi-agent pipeline architecture comprising a query agent, a planning agent, an execution agent, and a reporting agent.
50%