Agent Runtime Overview

Agent Runtime Overview

What is Agent Runtime?

Agent Runtime is a general-purpose platform for launching arbitrary AI agent workloads on demand. It orchestrates AI agents (Claude via Agent SDK, OpenCode) in isolated Modal sandboxes, providing a Node.js backend runtime and React client library for building applications with AI agents.

Key principle: The runtime handles session management and agent orchestration - it does not enforce any data schema on the calling application. Data types are passed back to the app, which decides what to do with them.

Packages

The monorepo provides two packages:

Why Use Agent Runtime?

Isolation

Agents execute in Modal sandboxes, not on the application server. This provides:

• Security isolation for arbitrary code execution
• Resource isolation (CPU, memory, disk)
• Clean environment for each session

Streaming

Real-time block-by-block streaming of agent output via WebSocket:

• block_start - New block begins
• text_delta - Incremental text updates
• block_update - Block metadata changes
• block_complete - Block finishes

Session Management

Built-in session lifecycle with:

• Lazy sandbox creation (sandbox spins up on first message, not session create)
• Idle timeout cleanup
• Periodic state sync to persistence
• Graceful shutdown

Multi-Architecture Support

Supports multiple agent architectures:

• Claude Agent SDK - Anthropic's official agent SDK
• OpenCode - Alternative agent runtime

Configure via AGENT_ARCHITECTURE_TYPE when creating sessions.

React-Ready

First-class React integration with hooks for:

• Session lifecycle management
• Message sending and streaming
• File workspace tracking
• Subagent transcript access

Architecture

┌─────────────────────────────────────────────────────────────┐
│                      Client Application                      │
│  ┌─────────────────────────────────────────────────────┐    │
│  │           AgentServiceProvider (React)               │    │
│  │  ┌──────────────┐  ┌──────────────┐  ┌───────────┐  │    │
│  │  │useAgentSession│  │ useMessages  │  │useFiles   │  │    │
│  │  └──────────────┘  └──────────────┘  └───────────┘  │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────┘
                    │ REST API        │ WebSocket
                    ▼                 ▼
┌─────────────────────────────────────────────────────────────┐
│                     Backend Runtime                          │
│  ┌─────────────────────────────────────────────────────┐    │
│  │                  SessionManager                      │    │
│  │  ┌──────────────────────────────────────────────┐   │    │
│  │  │              AgentSession                     │   │    │
│  │  │  - Sandbox lifecycle                          │   │    │
│  │  │  - Transcript parsing                         │   │    │
│  │  │  - File watching                              │   │    │
│  │  │  - Periodic sync                              │   │    │
│  │  └──────────────────────────────────────────────┘   │    │
│  └─────────────────────────────────────────────────────┘    │
│                           │                                  │
│  ┌────────────────────────┴────────────────────────────┐    │
│  │              PersistenceAdapter                      │    │
│  │  (Implemented by your application)                   │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│                      Modal Sandbox                           │
│  ┌─────────────────────────────────────────────────────┐    │
│  │     Agent (Claude SDK or OpenCode)                   │    │
│  │  - Tools (Read, Write, Edit, Bash, Grep, Glob)      │    │
│  │  - Skills                                            │    │
│  │  - Subagents                                         │    │
│  │  - MCP Servers                                       │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────┘

Data Flow

1. Client creates session via REST API
2. Runtime creates AgentSession (sandbox is NOT created yet - lazy initialization)
3. Client joins WebSocket room for the session
4. First message triggers sandbox creation, then agent execution begins
5. Agent output streams as block events via WebSocket
6. Session state periodically syncs to PersistenceAdapter

Key Concepts

Sessions

A session represents a single agent conversation. Sessions are:

• Created via REST API
• Identified by unique session ID
• Associated with an agent profile
• Persisted via the application's PersistenceAdapter

Blocks

Agent output is parsed into typed blocks for rendering:

• UserMessageBlock - User input
• AssistantTextBlock - Agent text response
• ToolUseBlock - Tool invocation
• ToolResultBlock - Tool output
• ThinkingBlock - Agent reasoning (if exposed)
• SystemBlock - System messages
• SubagentBlock - Subagent invocation
• ErrorBlock - Error information

Agent Profiles

Configuration defining agent capabilities:

• System prompt and memory file (CLAUDE.md/AGENT.md)
• Available tools
• Skills with supporting files
• Subagents for delegation
• Commands for specific workflows
• MCP server integrations

PersistenceAdapter

The main integration point between the runtime and application storage. Applications implement this interface to:

• Store and retrieve sessions
• Save transcripts
• Track workspace files
• Manage agent profiles

When to Use Agent Runtime

Good fit:

• Building applications that need AI agents running in isolated environments
• Launching arbitrary agent workloads on demand
• Applications requiring real-time streaming of agent output
• Multi-tenant systems where agent isolation is important
• Building custom AI interfaces (chat UIs, IDEs, automation tools)

Consider alternatives if:

• Running a single, long-lived agent process
• No need for sandbox isolation
• Simple request/response patterns without streaming

Related Skills

• backend-setup - Setting up the Node.js backend runtime
• react-integration - Building React frontends with the client library
• agent-design - Configuring agent profiles and workflows