AI Agent
Agent threads, tools and streaming
git clone https://github.com/bijectionhq/agents.git
cd agents
git checkout 8598acb6eeab2b2f76ac99feb4abcc3e66602de9Bijection repository access is required. Request access.
Description
Agent threads, tools and streaming
Explore the package README for API concepts, configuration, and examples. Check the Bijection source for adaptation-specific setup.
This package has a pinned source revision in Bijection’s workspace. Source availability does not establish npm publication or runtime qualification.
Getting started
- Get the source at the revision above and read its package documentation.
- Check the required Bijection SDK version, dependencies, and external service credentials.
- Follow the package setup instructions and the guide to using components.
- Exercise the component in your deployment according to your project’s verification rules.
Open source
Maintained source: bijectionhq/agents. Based on convex/agent. License: Apache-2.0. Consult the repository for license terms, notices, and contributor attribution.
From the upstream publisher
Convex component for AI agents with persistent chat threads, tool calling, streaming, and RAG.
Explore the demo ↗Demo video
Upstream overview
From the upstream package documentation.
The AI Agent component provides a structured framework for building agentic AI workflows with persistent message threads, automatic conversation context, and vector search capabilities. It separates long-running AI operations from your UI while maintaining real-time reactivity through Convex's websocket streaming. The component handles thread management, message persistence, file storage integration, and includes built-in debugging tools and usage tracking.
Use cases
From the upstream package documentation.
- Building customer support chatbots that need to maintain conversation history across multiple sessions and agents
- Creating multi-step AI workflows where different agents handle specific tasks like research, analysis, and reporting
- Implementing collaborative AI assistants where human agents and AI agents work together in shared threads
- Developing AI-powered applications that require RAG capabilities with hybrid vector and text search across conversation history
- Building usage-metered AI services that need per-user, per-agent cost tracking and rate limiting
How it works
From the upstream package documentation.
The component centers around Agents, Threads, and Messages as core abstractions. Agents encapsulate LLM prompting logic, tools, and behavior, while Threads persist conversation history that can be shared across multiple users and agents. Messages automatically include conversation context through built-in hybrid vector/text search within threads.
Streaming is handled through Convex's websocket infrastructure, enabling real-time text and object streaming without HTTP streaming complications. The component integrates with Convex file storage for automatic file handling with reference counting, and supports both static and dynamic workflows for complex multi-agent operations.
Debugging capabilities include callback functions, an agent playground for prompt iteration, and dashboard inspection. Usage tracking provides granular metrics per-provider, per-model, and per-user, while rate limiting integration helps manage API costs and provider limits.
Upstream package FAQ
How does the AI Agent component handle message persistence?
The Convex AI Agent component automatically stores conversation history in your Convex database with full ACID guarantees. Message threads persist across user sessions and remain queryable, with built-in indexing for fast retrieval during agent reasoning.
Can AI agents react to real-time database changes?
Yes, agents built with the Convex AI Agent component can subscribe to database mutations and trigger workflows automatically. This enables agents that respond immediately to new data, user actions, or external events without polling.
What vector search capabilities are included?
The AI Agent component includes built-in vector search powered by Convex's vector database features. Agents can perform semantic search over documents, conversation history, or custom embeddings without managing separate vector infrastructure.
How do long-running agent workflows work?
The Convex AI Agent component runs workflows as background processes that don't block your UI. Agents can execute multi-step reasoning, make API calls, and update state over time while users receive real-time progress updates through Convex's reactive subscriptions.
Do I need to manage separate infrastructure for AI agents?
No, the Convex AI Agent component runs entirely within your existing Convex backend. Message storage, vector search, and workflow execution all use Convex's infrastructure, eliminating the need for additional databases or orchestration tools.
Use with agents and CLI
Give your coding agent the component’s source identity and setup guidance.
Help me integrate AI Agent with Bijection.
Read: https://bijection.com/components/agents/SKILL.md
Check SDK compatibility and dependency availability.
Follow this project's execution and verification rules.Upstream README.md
View repository ↗Documentation from convex/agent, recorded October 23, 2026. APIs and setup may differ in the Bijection adaptation.
AI Agents, built on Convex. Check out the docs here.
The Agent component manages threads and messages that agents can use in static or dynamic workflows.
- Agents provide an abstraction for using LLMs to represent units of use-case-specific prompting with associated models, prompts, Tool Calls, and behavior in relation to other Agents, functions, APIs, and more.
- Threads persist messages and can be shared by multiple users and agents (including human agents).
- Streaming text and objects using deltas over websockets so all clients stay in sync efficiently, without http streaming. Enables streaming from async functions.
- Conversation context is automatically included in each LLM call, including built-in hybrid vector/text search for messages in the thread and opt-in search for messages from other threads (for the same specified user).
- RAG techniques are supported for prompt augmentation from other sources, either up front in the prompt or as tool calls. Integrates with the RAG Component, or DIY.
- Workflows allow building multi-step operations that can span agents, users, durably and reliably.
- Files are supported in thread history with automatic saving to file storage and ref-counting.
- Debugging is enabled by callbacks, the agent playground where you can inspect all metadata and iterate on prompts and context settings, and inspection in the dashboard.
- Usage tracking is easy to set up, enabling usage attribution per-provider, per-model, per-user, per-agent, for billing & more.
- Rate limiting, powered by the Rate Limiter Component, helps control the rate at which users can interact with agents and keep you from exceeding your LLM provider's limits.
Read the associated Stack post here.
Read the docs for more details.
Play with the example:
git clone https://github.com/convex/agent.git
cd agent
npm run setup
npm run dev
Found a bug? Feature request? File it here.