Overview
Andrew describes how an internal attempt to relieve Vercel’s overloaded data team became a broader framework for building production agents. The experiment began with a Snowflake schema pasted into a large prompt and manually executed SQL, then progressed to a chain of specialized agents for planning, querying, execution, and reporting. Although this pipeline automated the full journey from question to answer, its compressed handoffs and prescribed tools limited reflection, recovery, and performance on real employee requests. A single stateful agent improved error correction, but trusted users still found it unreliable. The decisive breakthrough came from studying Claude Code and Opus 4.5: agents performed better when given a sandboxed filesystem and a small set of familiar capabilities such as reading files, writing files, and running Bash. Adding Vercel-specific tools and roughly 100 automatically distilled skills doubled evaluation performance and supported thousands of daily internal queries. Those lessons became Eve, a convention-based framework organized around skills, tools, channels, runtime services, and adapters. Andrew argues that generic vertical agents have limits because high-value automation depends on company-specific semantics and operating knowledge. Vercel now reportedly has about 20 agents with meaningful product-market fit, spanning data, marketing, sales, and legal work, allowing specialists to spend less time servicing repetitive requests and more time improving their underlying systems.
Sections
Strategic Insights
Higher-level implications derived from the architecture and adoption journey.
- The project’s progress came primarily from changing the agent’s operating environment, not from continually subdividing reasoning into more agents. The filesystem, sandbox, and familiar shell operations gave the model a flexible workspace in which it could externalize state and recover from errors.
- D0 evolved from stateless generation toward a learning system at two levels: individual runs preserved working state, while recurring query patterns were distilled into reusable skills across runs. This combination turned both immediate experience and historical usage into context.
- The poor reception from trusted users shows that evaluation performance and practical usefulness were not equivalent. Real adoption supplied a broader request distribution and revealed that architecture optimized around anticipated scenarios could remain brittle.
- Eve’s framework thesis mirrors Vercel’s web-platform strategy: developers declare intent through filesystem conventions while infrastructure concerns are handled beneath the application surface. The company is transferring an established product philosophy from web deployment to agent deployment.
- The strongest internal return described is organizational leverage. Automating request fulfillment allows expert teams to improve the systems that generate future leverage, creating a potentially compounding productivity effect.
Technical Architecture
Concrete implementation details mentioned in the talk.
- The AI SDK provides a shared model interface across providers, reducing a provider migration from hundreds of provider-specific lines to approximately one changed line.
- The earliest data-agent prototype embedded a Snowflake schema and user question in a system prompt, generated SQL, and relied on Andrew to execute that SQL manually.
- The multi-agent D0 architecture separated query intake, planning, SQL execution, and reporting. Individual agents had focused prompts and scoped tools such as reading entity YAML files and searching schemas.
- The consolidated architecture used one large agent call with a maximum of approximately 100 steps, allowing the same agent to plan, build, execute, report, and revisit earlier work.
- The filesystem architecture runs inside a sandbox populated with the semantic layer. Its core interface includes Bash plus file listing, reading, and writing, supplemented by a small number of Vercel-specific tools.
- A Bash helper available through npm can be attached to a sandbox, and files can be mounted for the agent to read, modify, and execute.
- A recurring job analyzes recent queries and distills common patterns into a skills directory; the system currently contains roughly 100 skills.
- Eve models an agent as a runtime plus channels. Runtime services include durability, isolated execution, model invocation, and connections, while filesystem conventions expose system instructions, skills, and tools.
- Eve supports open-source adapters for PostgreSQL, OpenAI’s Responses API, Docker, and other connectors. Its Vercel deployment path integrates Workflows for durability, Sandbox for secure execution, and Vercel Connect for short-lived OIDC connection tokens.
- Vercel deployment includes observability for agent runs, tool calls, execution steps, estimated costs, and possible optimizations.
Evolution from Prototype to Eve
The chronological development of Vercel’s internal agent architecture.
- Andrew and Vercel’s CTO proposed the idea of placing an agent on every desk, then interviewed teams across marketing, sales, finance, and legal to identify painful work.
- The first prototype generated SQL from a Snowflake schema embedded in a large prompt, with manual execution used to test basic model capability.
- D0 was decomposed into specialized agents responsible for querying, planning, execution, and reporting.
- The team replaced the chained architecture with one stateful agent capable of managing planning, execution, reporting, and error recovery within a long run.
- A trusted internal release received strongly negative feedback despite the system passing roughly 30% of its evaluations.
- After Claude Code and Opus 4.5 appeared, the team rebuilt D0 around a sandboxed filesystem and a minimal set of familiar shell and file tools.
- Broader internal adoption grew to thousands of daily queries, and recurring jobs began turning repeated query patterns into roughly 100 reusable skills.
- Vercel rewrote D0 in Eve, tested the framework with beta customers including Aura, and released it publicly at a London event.
Forecasts and Forward-Looking Claims
Expectations Andrew expresses about the future of business automation and agent development.
- Agents will expand beyond coding and technical work into design, product management, and other professional verticals.
- More organizations will build business-specific agents rather than relying exclusively on generic, off-the-shelf vertical products.
- Companies of all sizes will increasingly automate repetitive work in functions such as HR, finance, and sales.
- Company-specific knowledge will remain central to achieving strong performance because generic agents lack the internal semantics needed to interpret proprietary data and workflows.
- Convention-based frameworks such as Eve will let teams start from mature agent-building patterns instead of repeating the progression from simple prompts through increasingly complex custom orchestration.