𝗔𝗴𝗲𝗻𝘁 𝘄𝗶𝘁𝗵 𝗩𝗲𝗿𝗰𝗲𝗹'𝘀 𝗘𝘃𝗲 𝗙𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸
Vercel open-sourced Eve. It is a framework to build durable AI agents.
Instead of writing complex code for model loops or tool dispatch, you write files in a directory. Eve handles the rest.
I built a shopping assistant using Eve. It searches products, checks inventory, compares prices, and places orders.
Here is what I learned:
The Core Concept Eve separates the agent from the communication channel.
- The agent is the reasoning core. It contains the model, tools, and instructions.
- The channel is the delivery method. It handles auth, message formats, and platforms.
This design lets one agent serve many surfaces at once. You can use the same agent for a web chat, Slack, a CLI, or a webhook. You add new surfaces by adding channel files, not by changing your agent code.
Durable Sessions Eve uses a workflow engine to keep conversations alive.
- Sessions survive server restarts.
- If a process crashes mid-turn, Eve resumes from the last completed step.
- You can stop your server, restart it, and continue the same conversation.
For local development, Eve saves everything to your filesystem in a .workflow-data directory. For production, you can plug in different backends like PostgreSQL, Redis, or MongoDB.
Zero Orchestration I did not write any routing or streaming logic.
- The filesystem handles the setup. To add a capability, you just create a file.
- The agent handles multi-step reasoning naturally.
- The developer tool provides a real-time terminal UI to see tool calls and streaming text.
If you use a custom model provider, remember to define your modelContextWindowTokens. Eve uses this to summarize old messages so your conversation never hits a limit.
Key Takeaways:
- No boilerplate code for tool dispatch.
- Built-in session persistence.
- Easy multi-platform support via channel files.
- Fast development with a real-time TUI.
Source: https://dev.to/stormhub/agent-with-vercels-eve-framework-3c2l
Optional learning community: https://t.me/GyaanSetuAi