get-convex/agent

What it solves

This project provides a framework for building AI agents that can maintain persistent chat threads, handle tool calling, and integrate RAG (Retrieval-Augmented Generation). It solves the problem of keeping long-running agent workflows separate from the UI while ensuring clients receive reactive updates in real-time.

How it works

The component manages threads and messages that agents use in either static or dynamic workflows. It uses an abstraction for LLMs to handle specific prompting, models, and tool calls. To keep clients in sync, it streams text and objects via websockets rather than traditional HTTP streaming. It automatically includes conversation context, including hybrid vector/text search for messages within a thread, and supports RAG techniques for prompt augmentation from external sources.

Who it’s for

Developers building AI-powered applications on the Convex platform who need a durable, reliable way to manage agentic workflows, conversation history, and real-time updates.

Highlights

• Persistent Threads: Manages shared messages between users and agents, including human agents.
• Reactive Streaming: Uses websockets for efficient streaming of text and objects from async functions.
• Integrated RAG: Supports hybrid vector/text search and integrates with the RAG Component for prompt augmentation.
• Durable Workflows: Enables multi-step operations that span across agents and users.
• Developer Tooling: Includes an agent playground for prompt iteration, usage tracking for billing, and rate limiting to manage LLM provider limits.