TechPrufer
Get started
Vercel
Vercel logoVercel · Feature trail

Vercel Workflows

15 posts · Feb 2023 – Aug 2026

What this trail covers

Vercel Workflows provides a programming model for durable execution of long-running systems like agents, backends, and ETL pipelines. It integrates with the AI SDK for durable agents and provides a Python SDK beta. The system eliminates separate orchestrators by running coordination within application code, leveraging an event log, Fluid compute for function execution, and Vercel Queues. Security is enhanced with default encryption of all data. The Community Guardian is now generally available. Eve, Vercel's open-source agent framework, simplifies agent development by providing pre-built scaffolding for channels, connections, tools, approval gates, and durable workflows, allowing developers to focus on the agent's core logic.

All Vercel trails
  1. The best workflow engine is a programming languageThis post introduces the Workflow SDK, an open-source library that enables writing durable workflows in TypeScript using familiar programming language constructs. It replaces traditional workflow engines by treating code itself as a Directed Acyclic Graph (DAG). Key contributions include: 1. A 'use workflow' and 'use step' directive system for defining workflow orchestrators and side-effecting units of work. 2. Automatic retries for steps with configurable error handling (FatalError, RetryableError). 3. A unified 'hook' primitive that replaces Temporal's signals, queries, and updates, simplifying inter-workflow communication and external interaction via ad-hoc webhooks. 4. A swappable backend architecture ('World') allowing integration with various storage (Postgres, Turso, Durable Objects) and queuing systems (Vercel Queues, SQS). 5. A Vercel-specific versioning strategy that pins runs to the starting deployment, simplifying in-flight workflow evolution. 6. A stated goal of minimizing step overhead to make them feel 'free'.
  2. eve: AI agents without the plumbingThis talk introduces Eve, Vercel's open-source agent framework, which significantly reduces the 'plumbing' required to build and deploy AI agents. Eve provides pre-built components for channels (Slack, Discord, etc.), secure credential management (MCP/OpenAPI), tools (TypeScript in Vercel Sandbox), approval gates, and durable workflows with checkpointing and crash recovery. The framework allows agents to be defined as a simple directory with an instruction file, abstracting away the complexities of infrastructure, authentication, and state management, enabling developers to focus on agent logic and ship hundreds of agents in production.
  3. The Agent StackThis post introduces the 'Agent Stack' as a cohesive set of building blocks for creating production-grade AI agents. It details three core capabilities: connecting to models (AI SDK, AI Gateway), executing complex workflows (Workflow SDK, Vercel Sandbox), and connecting to data and tools (Vercel Connect, Chat SDK). It highlights the tradeoffs developers face and how the Agent Stack addresses them, emphasizing durability, security, and ease of integration. The 'eve' framework is presented as a structural pattern for agent development.
  4. Introducing eveIntroduced 'eve', an open-source agent framework designed to standardize agent development by providing built-in production plumbing such as durable execution (via Workflow SDK), sandboxed compute (via Vercel Sandbox), human-in-the-loop approvals, subagents, evals, and secure connections to tools and services (via Vercel Connect). eve simplifies agent creation by defining components like models, instructions, tools, skills, subagents, channels, and schedules in a directory structure, reducing boilerplate and enabling faster iteration.
  5. A new programming model for durable executionThis post announces the general availability of Vercel Workflows, highlighting its expansion to Python with a beta SDK and deep integration with the AI SDK for durable agents. It details the core components (event log, Fluid compute, Vercel Queues) and the programming model where application code acts as the orchestrator. It also emphasizes security features like default encryption and introduces the concept of durable streams for agent output.
  7. Keeping community human while scaling with agentsThis post details the implementation of the Community Guardian, an agent-based system for managing Vercel's developer community interactions. It highlights the use of Vercel Workflows for the operational layer, enabling durable execution of agent tasks such as post analysis, routing, duplicate detection, and follow-up reminders. The agent 'c0' is described as the intelligence layer, using AI Gateway and AI SDK to research community threads and extract product feedback. The post emphasizes how this system allows human community managers to focus on higher-value interactions by automating repetitive tasks. It also provides insights into building similar agent systems using Vercel's AI tools and the Chat SDK.
  8. Anyone can build agents, but it takes a platform to run themThis post details how Vercel uses its own platform primitives (Sandboxes, Fluid Compute, AI Gateway, Vercel Workflows) to run internal AI agents like d0 (an analytics agent). It highlights the shift from 'build vs. buy' to 'build and run' for AI agents, emphasizing the operational complexity of production deployment. The post explains the architecture of d0, including its interaction with Slack, a semantic layer, the AI SDK, Vercel Workflows for durable orchestration, Vercel Sandboxes for isolated execution, AI Gateway for model routing, and the AI SDK Chatbot primitive for delivering results. It also lists other internal and product agents running on Vercel's platform, underscoring the role of Sandboxes for security, Fluid Compute for cost-effective scaling, and AI Gateway for model flexibility and resilience.
  9. Testing if "bash is all you need"This post details an evaluation of different agentic approaches for querying structured data, comparing SQL, bash (using just-bash), and filesystem-based agents. It highlights performance bottlenecks in bash (e.g., stat() calls), the need for schema context, and eval scoring issues. The key contribution is the demonstration of a hybrid approach combining bash for exploration/verification with SQL for direct querying, achieving 100% accuracy with a trade-off in cost. It also emphasizes the importance of robust evaluation infrastructure and iterative debugging for improving agent capabilities.
  10. How Mux shipped durable video workflows with their @mux/ai SDKThis post details how Mux integrated Vercel's Workflow DevKit into their `@mux/ai` SDK to enable durable AI video workflows. It highlights the problem of AI workflow failures and how Workflow DevKit's "use workflow" and "use step" directives provide opt-in durability without hard infrastructure requirements or new DSLs. The post demonstrates a practical example of a durable video AI workflow and explains how Vercel handles state storage and observability automatically. It also mentions the availability of pre-built workflows for common video AI tasks and the open-source nature of the SDK.
  11. How to build agents with filesystems and bashThis post introduces a new pattern for agent context management by leveraging filesystems and bash tools. It argues that LLMs are already proficient with these tools from their training on codebases, and therefore, structuring agent data as files and providing bash access allows agents to explore and retrieve context more effectively than traditional prompt stuffing or vector search. The post details how to map domains to filesystem structures with examples for customer support and document analysis, and presents a case study of a sales call summary agent that saw cost reductions and quality improvements. It highlights the benefits of this approach, including native model capabilities, future-proofing, debuggability, security through isolation, and reduced maintenance code.
  12. Inside Workflow DevKit: How framework integrations workThis post details the pattern behind Vercel Workflow DevKit (WDK) framework integrations. It explains the two-phase pattern: build-time generation of workflow handlers using SWC compiler plugins (client, step, and workflow modes) and runtime exposure of these handlers as HTTP endpoints via framework-specific routing. It highlights how WDK adapts to different frameworks like SvelteKit (Vite-based) and Express/Hono (HTTP server-based, using Nitro), including handling framework-specific request object conversions and implementing Hot Module Replacement.
  13. Workflow Builder: Build your own workflow automation platformIntroduces Workflow Builder, an open-source Next.js template for building workflow automation platforms. It showcases the use of the Workflow Development Kit (WDK) for creating visual workflow editors, AI-powered text-to-workflow generation, and webhook triggers. The post details how these components compile into executable TypeScript, enabling the development of agents, internal tools, and customer-facing workflow solutions. It highlights integrations with Resend, Linear, Slack, PostgreSQL, HTTP requests, and Vercel AI Gateway.
  14. Vercel Ship AI 2025 recapAI SDK 6 introduces an agent abstraction with tool execution approval for human-in-the-loop review, allowing agents to pause for confirmation before executing actions. The Vercel Marketplace now features agents and AI services. The Workflow Development Kit (`use workflow`) adds reliability-as-code to TypeScript functions, enabling automatic retries, progress persistence, and built-in observability. Vercel Agent offers AI code reviews with validated suggestions and anomaly alerts that trigger automated investigations. Python support with zero configuration is now available via the Vercel Python SDK for frameworks like FastAPI and Flask. Open-source lead agent and data analyst agent templates are provided.
  15. Built-in durability: Introducing Workflow Development KitIntroduces the Workflow Development Kit (WDK), an open-source TypeScript framework that makes durability a language-level concept. WDK adds `use workflow` and `use step` directives to async functions, enabling them to pause, persist, and resume across deployments and crashes. It automatically handles queues, retries, and persistence, allowing workflows to pause for extended periods (minutes to months) and resume exactly where they left off. WDK supports pausing with webhooks for external events and is designed to run on any platform via 'Worlds', including a local development environment and Vercel's production environment using Framework-defined infrastructure. The post details the core concepts with code examples for ride-hailing and loyalty reward workflows, and a payment validation workflow using webhooks. It emphasizes that WDK simplifies building reliable systems by abstracting away complex distributed systems infrastructure.
  16. You can just ship agentsThis post introduces the Workflow Development Kit (WDK) which allows developers to express durability directly in TypeScript for long-running agentic tasks. It highlights how the WDK enables agents to plan and act over time, persisting progress at each step and resuming after crashes or deployments. The WDK integrates with Vercel's platform for execution and provides traces for observability, simplifying the development of resilient agentic systems.
  17. Agents at work, a partnership with Salesforce and SlackThis post details a partnership between Vercel and Salesforce to integrate Vercel's AI Cloud capabilities with Slack, enabling AI agents to operate within conversational interfaces. It highlights the concept of agents living in Slack threads, interpreting intent, and taking action. The post introduces the v0 Slack Agent as a demonstration of collaborative agent development and deployment directly from Slack. It emphasizes the Vercel AI Cloud as the runtime for these agents, supporting long-running tasks, stateful sessions, and secure tool execution. The partnership aims to bring Vercel's agent infrastructure to where teams collaborate, extending frontend extensibility to frameworks like Next.js and Agentforce 360.
  18. The second wave of MCP: Building for LLMs, not developersThis post introduces the concept of "intention-based tools" for LLMs, contrasting them with "API-shaped tools." It explains that LLMs lack the state management capabilities of developers, requiring tools that encapsulate complete user intentions (workflows) rather than exposing individual API operations. The post provides examples of how a `deploy_project` tool is more effective for LLMs than a sequence of `create_project`, `add_env`, `deploy`, and `add_domain` calls. It advocates for testing workflows manually and designing tools around user goals, using regular programming for deterministic parts and LLMs for reasoning. The goal is to achieve first-try success for complex workflows and improve reliability and efficiency.
  19. Model Context Protocol (MCP) explained: An FAQThis post introduces the Model Context Protocol (MCP) as a specification for standardizing how Large Language Models (LLMs) access external data and systems. It explains MCP's purpose as a universal connector for AI agents, enabling them to interact safely and predictably with applications by exposing data sources and tools. The post differentiates MCP servers from traditional APIs, highlighting that MCP is designed for models and provides metadata for tool discovery and usage. It details the concept of 'transports' (e.g., StreamableHTTP, stdio) for communication between models and MCP servers, and distinguishes between local and remote MCP server deployments. Use cases are provided for e-commerce, finance, and marketing, emphasizing how MCP allows models to take concrete actions. The post also mentions Vercel's AI SDK and `mcp-handler` as tools for creating MCP servers, and clarifies the relationship between AI agents and MCP servers: agents plan actions, while MCP servers define the available actions.
  20. The no-nonsense approach to AI agent developmentThis post outlines a practical, step-by-step approach to building AI agents, emphasizing a 'no-nonsense' methodology. It advocates for manual prototyping of the agent's task before coding, followed by automating the loop with deterministic code for non-reasoning parts and LLMs for judgment. The final step focuses on optimizing for reliability through prompt refinement, precise tool calls, and structured evaluations. The core message is that building reliable AI agents leverages fundamental programming principles rather than requiring entirely new abstractions.
  21. Introducing Vercel Cron JobsIntroduces Vercel Cron Jobs, a feature for scheduling Vercel Functions. It details configuration via `vercel.json` with cron expressions, provides examples of schedules, and outlines limits for different plans. It also mentions the potential for framework-native integration via the Build Output API.