Vercel eve
3 min read
Vercel eve is an open-source, filesystem-first TypeScript framework for building and running backend agents. An agent is defined as a directory containing instructions, tools, skills, connections, channels, sandbox configuration, subagents, and schedules. The framework owns the runtime plumbing around those files.
Vercel introduced eve publicly in June 2026. It is in public preview, so its interfaces and platform assumptions should be treated as active rather than settled.
What eve owns
Section titled “What eve owns”eve combines several responsibilities that otherwise require separate infrastructure:
| Capability | eve’s approach | Related concept |
|---|---|---|
| Agent definition | Markdown instructions and skills, TypeScript tools and configuration | Agent, Skills |
| Durable sessions | Each conversation runs as a checkpointed durable workflow | Durable Agent Execution, Orchestrator |
| Human approval | Tool actions can pause until a person approves | Controls & Autonomy |
| Execution isolation | Agent-generated code runs in a separate sandbox | Security & Prompt Injection, Agent Architecture |
| Channels | Adapters connect the same agent to HTTP, chat, issue trackers, and other surfaces | Agent-Computer Interface |
| Connections | MCP and OpenAPI-backed services are discovered with brokered credentials | MCP |
| Tracing | Model and tool calls emit OpenTelemetry spans | Execution Ledger |
| Evaluation | File-based evals can run locally, remotely, or in CI | The Evaluation Harness |
The attraction is integration. Durable execution, approvals, channels, tracing, and evals share one session model instead of being assembled as unrelated subsystems.
Durability model
Section titled “Durability model”eve treats every conversation as a durable workflow. Each step is checkpointed, allowing a session to survive a crash or deployment and resume from its previous position. Approval-gated actions can suspend without holding compute and continue when the approval arrives. This is one implementation of the broader Durable Agent Execution pattern.
This directly supports asynchronous workflows that span hours or days. It also aligns with the book’s rule that nondeterministic work must be isolated from replayable orchestration. Model and tool results belong in durable history so recovery does not silently produce a different decision or repeat a completed side effect.
The framework provides durability mechanics, not complete business correctness. A production design must still define idempotency for writes, stale-approval behavior, authorization at resume time, compensation for partial failure, and versioning for workflows already in flight.
Channels, sessions, and users
Section titled “Channels, sessions, and users”eve separates an agent from the surfaces where people use it. A channel adapter normalizes platform events into session turns, owns the continuation token, and returns output through the originating surface. The same agent can therefore appear in a web interface, Slack, GitHub, Linear, or a custom channel without duplicating its core behavior.
Channels also form an identity boundary. User-scoped connections can resolve credentials for the current participant, allowing a tool to act with that person’s access rather than one shared agent credential. Shared threads still require explicit rules for who may resume a session or approve an action.
Traces and the learning loop
Section titled “Traces and the learning loop”Every eve run produces an ordered trace of turns, model calls, and tool calls using OpenTelemetry. Those traces cover the agent side of the Self-Learning Agents loop.
They do not automatically capture every meaningful correction made elsewhere in the application. The product still needs to instrument user outcomes, correlate them with the eve session, and write the joined record into its evaluation and learning pipeline.