Vercel Eve Agent-as-Directory Workflow

Vercel Eve Agent-as-Directory Workflow is a filesystem-first approach to building production AI agents. Instead of writing registration boilerplate, defining graph state, or managing deployment infrastructure separately, you create a single directory with Markdown files for prompts and TypeScript files for tools. The framework discovers these files at build time, compiles a manifest, and runs the agent as a durable, checkpointed workflow on Vercel Functions. Agentic reasoning happens when the LLM reads the instructions.md system prompt and decides which tool to call based on the current conversation state. The outcome is measurable: Vercel runs over 100 internal agents on Eve including a Slack analyst handling 30,000 questions per month and a support agent resolving 92% of tickets without human intervention. (Source: Vercel Blog, Introducing Eve, June 17, 2026; Byteiota, Vercel Eve Framework, June 18, 2026.) A single developer can scaffold, add three tools, and deploy a production agent in under 30 minutes. The agent survives process crashes, deploys, and restarts because the durable runtime checkpoints state between every step. Teams that adopt the workflow report agent onboarding time dropping from 3 days to 2 hours per agent because the directory convention replaces custom scaffolding. The cost to start is zero: Eve is Apache 2.0 and the npx init command requires only Node.js 20+ and npm.

BUSINESS PROBLEM

Teams building AI agents today repeat the same non-differentiating infrastructure work for every agent. A 10-person engineering team at a Series B SaaS company building three agents per quarter spends 9 to 15 developer-days per quarter wiring up tool registration, state management, credential rotation, sandboxed execution, and deployment pipelines before writing a single line of agent-specific logic. At a blended rate of $800 per developer-day, that scaffolding tax costs $7,200 to $12,000 per quarter. LangGraph requires assembling Tool, Chain, and Memory abstractions in Python. CrewAI demands role definitions and process graphs. Mastra provides a more structured approach but still requires manual tool registration and state persistence configuration. None of these frameworks ship with built-in durable execution, credential management, or sandboxed compute. Vercel CEO Guillermo Rauch stated at Ship London that agents now trigger more than half of all commits on Vercel platform, up from under three percent six months ago. (Source: Tech Times, Vercel Eve Launch Coverage, June 18, 2026.) As agent quantity grows, the scaffolding burden multiplies linearly. Early adopters of Eve report that the directory convention eliminates 80 percent of the per-agent boilerplate. The opportunity is to adopt a convention that makes agent infrastructure a solved problem so teams focus on agent behavior, not agent plumbing.

WHO BENEFITS

Profile 1: Full-stack developer shipping a single-agent SaaS feature. ROLE: Frontend or backend developer at a 5 to 20 person product team. SITUATION: You need to deploy an AI agent that answers customer questions using product data. You do not want to learn graph abstractions or manage a separate state database. PAYOFF: Eve scaffolds a working agent in 2 minutes. Add three tool files using defineTool and Zod schemas, run vercel deploy, and the agent runs with durable execution and sandboxed compute on day one.

Profile 2: Platform engineer at a 20 to 100 person product team. ROLE: Infrastructure or platform team member responsible for supporting multiple agents across product squads. SITUATION: Your team maintains 5 to 15 agents, each built with a different framework and each with custom credential, deployment, and observability plumbing. PAYOFF: Eve imposes one directory convention across all agents. Tool files become PR-reviewable diffs. Vercel Connect replaces manual token rotation. First month: agent onboarding drops from 3 days to 2 hours.

Profile 3: Startup CTO evaluating agent frameworks. ROLE: CTO or engineering lead at a seed to Series A startup. SITUATION: You are choosing between LangGraph, CrewAI, Mastra, and Eve for an AI product feature that must ship in 4 weeks. PAYOFF: Eve Apache 2.0 license and Vercel deployment pipeline mean the team writes zero infrastructure code. First agent ships in week 1 with built-in durability, sandboxing, and observability.

HOW IT WORKS

Step 1 — Scaffold the Agent · Tool: Vercel Eve 0.1.x + Node.js 20+ · Time: 2 minutes

Input npx eve@latest init customer-support-agent Action The CLI installs the eve npm package, creates the agent/ directory with default agent.ts and instructions.md, initializes Git, and starts a dev server. Output A working agent directory at customer-support-agent/agent/ with agent.ts, instructions.md, and the default model configuration.

Step 2 — Write the System Prompt · Tool: Any text editor · Time: 10 minutes

Input Markdown text defining the agent role, rules, tone, and output format Action Eve reads instructions.md at build time and prepends it as the system prompt to every model call. The AI reasoning step evaluates user input against the instructions to decide intent. Output agent/instructions.md with 20 to 80 lines of Markdown defining the agent identity and behavioral guardrails.

Step 3 — Configure the Model and Runtime · Tool: Vercel Eve + AI Gateway · Time: 5 minutes

Input defineAgent({ model: "anthropic/claude-opus-4.8" }) in agent.ts Action Eve validates the model string against the AI Gateway provider registry. The model field supports provider/model format with automatic fallback configuration. If agent.ts is omitted, Eve uses its default model. Output agent/agent.ts with model selection and optional runtime fields like maxTokens, temperature, and compaction settings.

Step 4 — Add Typed Tools · Tool: TypeScript 5.x + Zod · Time: 15 minutes per tool

Input A file at agent/tools/{tool-name}.ts exporting defineTool with inputSchema (Zod) and an execute function Action Eve discovers the file at build time. The filename becomes the tool name exposed to the model. The Zod schema validates all incoming arguments before execute runs. File-based discovery means adding a capability is a one-file change reviewable in a PR. Output agent/tools/ directory with one .ts file per tool, each with typed input validation and an async execute handler. The sandbox context provides isolated filesystem and env access.

Step 5 — Add Skills as Markdown · Tool: Any text editor · Time: 10 minutes

Input Markdown files in agent/skills/ describing procedures, policies, or domain knowledge Action Eve embeds each skill file and evaluates relevance against the conversation state using embedding similarity. Only the matching skill is injected into the context window, preventing prompt bloat. Output agent/skills/ directory with .md files loaded on demand when the agent detects a relevant query based on semantic similarity.

Step 6 — Connect Channels · Tool: Vercel Eve + Vercel Connect · Time: 15 minutes

Input A channel adapter file at agent/channels/{channel-name}.ts Action Eve registers the channel at build time. The same agent logic runs across all connected channels without modification. Vercel Connect handles OAuth token exchange and short-lived credential rotation. Output agent/channels/ with one adapter file per surface (slack.ts, discord.ts, web.ts). The agent responds in all channels from a single runtime instance.

Step 7 — Deploy to Production · Tool: Vercel CLI · Time: 5 minutes

Input vercel deploy from the agent project root Action The Vercel CLI builds the Eve project, switches the sandbox adapter from Docker (local) to Vercel Sandbox (production), assigns a production URL, and provisions the durable workflow runtime. Output A production HTTPS endpoint with durable execution, OpenTelemetry tracing, and sandboxed compute. No infrastructure configuration required.

Step 8 — Add Human-in-the-Loop Approval · Tool: Vercel Eve needsApproval · Time: 5 minutes

Input Add needsApproval: true to any tool definition Action Eve pauses execution before executing the tool, persists the pending state to the durable store, and sends a notification through the configured channel. The agent resumes from persisted state when a human approves, even if the process restarted. Output A paused agent state that survives deploys and restarts. Approval arrives through Slack or email and the agent continues from the exact pending step.

TOOL INTEGRATION

TOOL: Vercel Eve 0.1.x Role: The agent framework that discovers files, compiles the manifest, and runs the durable runtime

ROI METRICS

Metric | Before | After | Source ---|---|---|--- Agent scaffolding time | 3-5 days per agent | 2 minutes (npx eve@latest init) | Vercel, Eve Docs, June 2026 Tool registration per agent | 45-90 minutes manual wiring | 0 minutes (file discovery) | Vercel, Introducing Eve Blog, June 2026 Sandbox setup per agent | 2-4 hours Docker config | 0 minutes (built-in adapter) | Vercel, Sandbox Docs, June 2026 Credential management overhead | 30 min/week rotating tokens | 0 min (Connect auto-rotation) | Vercel, Connect Blog, June 2026 Agent-triggered deploys share | 3% of all deploys | 50%+ of all deploys | Ship London Keynote, June 2026 Internal agent count at Vercel | N/A (fragmented) | 100+ agents on Eve | Vercel Blog, June 2026

CAVEATS

1. (high severity) Agent crashes with unhandled errors in tool execution produce no automatic retry. Eve durable checkpointing persists state between steps, but if a tool execute function throws an exception, the session terminates. Mitigation: wrap every tool execute body in try-catch and return structured error objects instead of throwing.

2. (moderate severity) Vercel Sandbox has a 10-second execution timeout per tool call. Tools that call external APIs with slow responses time out. Mitigation: design tools to return partial results within 10 seconds or delegate long-running work to a subagent with its own sandbox.

3. (moderate severity) Skills relevance matching uses embedding similarity with no user-exposed tuning. Similar skill language may load the wrong skill or load both, increasing token usage. Mitigation: write skill filenames and opening paragraphs with distinctive, narrow keywords. Test with eve eval before deploying.

4. (minor severity) Eve Connect integrations require the end user to have a Vercel account with correct team permissions. Customers self-serving the agent through an embedded UI must authenticate with Vercel first. Mitigation: use API-key-based tools for customer-facing agents and reserve Connect for internal tools. Vercel Passport solves this for enterprise.