Eve vs Building from Scratch

Eve is a convention-over-configuration TypeScript agent framework. An agent is a directory, and each subfolder maps to a specific capability: tools/ for callable functions, skills/ for reusable prompts, subagents/ for hand-offs, channels/ for input surfaces (Slack, email, HTTP), schedules/ for cron jobs. Vercel calls this "Next.js for agents," and the analogy holds — you get filesystem-based routing but for agent capabilities instead of pages.

Under the hood it composes three Vercel primitives: the Workflow SDK (durable, checkpointed execution), Vercel Sandbox (isolated micro-VMs for running untrusted or LLM-generated code), and AI Gateway (unified provider routing, observability, BYOK). This is the real value — three services that agent apps typically need to wire together are one install away.

A directory-shaped agent is a folder you scan at startup. Read tools/*.ts, import each, build a dict. Read subagents/*.ts, wrap each as a function. That's ~30 lines of Node.js filesystem code.

Durable execution is the harder piece. The naive version: serialize agent state (messages, tool history) to JSON or Postgres after each step, load on restart. Fine for most cases. Vercel Workflow SDK does this cleanly with automatic replay, which is genuinely useful if your agents run for hours. Sandboxed exec is child_process.spawn in a container, or a call to a Sandbox-as-a-service. Model routing is a switch statement across providers.

The full pattern — tool dispatch, sub-agent hand-off, checkpointing, cron schedules — fits in about 80 lines of TypeScript if you don't need micro-VM isolation. Eve's runtime pieces are what you're paying for, not the conventions.

Eve is the right pick when your agent needs durable execution, sandboxed code exec, and provider portability, and you're OK deploying on Vercel (or willing to run its runtime pieces yourself). Long-running agent workflows that need to survive crashes, LLM-generated code that shouldn't run in your main process, and multi-model routing all come for free.

If your team already ships on Vercel and uses Next.js, the mental model is instantly familiar. The convention-based structure also pays off on multi-agent codebases where you want a consistent way for engineers to add new tools or sub-agents without inventing conventions per project.

If your agent is a single tool-using loop with no crash-recovery needs, Eve is over-engineered. You don't need a Workflow SDK to append to an array. You don't need a Sandbox to call fetch. A 60-line fetch-based agent handles most production workloads and stays legible.

Eve's abstractions pay off at the workflow layer — durable long-running agents, sandboxed exec, sub-agent hand-offs. Below that layer, the conventions add cognitive overhead without saving code. Start with a function and a loop; adopt Eve when durability or sandboxing becomes the actual bottleneck.