AgentOS

Note: For detailed configuration, see AgentOS Configuration.

Overview

AgentOS is a V8 isolate + WASM-based virtual operating system by Rivet. It runs inside each Rivet actor and provides a POSIX-compatible environment for the Pi Agent instance.

Package: @rivet-dev/agentos (npm)
Docs: agentos-sdk.dev
GitHub: rivet-dev/rivet
Tagline: “Give agents an operating system as a library”

Role in Architecture

graph TB
    subgraph "Rivet Actor"
        AOS[AgentOS<br/>V8 Isolate + WASM]
        PI[Pi Agent Core<br/>Python runtime]
        TOOLS[Extensions]
    end

    subgraph "Storage"
        S3[(S3 Mount)]
        MEM[(Memory Mount)]
    end

    AOS --> PI
    PI --> TOOLS
    AOS -.->|mounts| S3
    AOS -.->|mounts| MEM

    style AOS fill:#fff9c4,stroke:#f9a825,stroke-width:3px
    style PI fill:#e1f5fe,stroke:#0277bd,stroke-width:2px

Key Characteristics

Characteristic	Value
Isolation model	V8 isolate + WASM
Memory per isolate	~22 MB
Cold start	4.8 ms p50
POSIX compatibility	Filesystem, network, processes, shell
Storage mounts	Host dir, S3, Google Drive, memory
Sandbox escalation	Can mount full Linux sandbox if needed

What It Provides

Virtual filesystem — mount S3, host dirs, or memory
Virtual networking — isolated outbound HTTP
Virtual processes — optional subprocess support
Virtual shell — command execution in sandbox
Resource limits — per-isolate CPU, memory, file, socket caps
Preview URLs — for debugging (optional)

Integration with Rivet

flowchart LR
    MSG[Incoming message] --> RIVET[Rivet get_actor]
    RIVET --> ACTOR[Rivet Actor]
    ACTOR --> AOS[AgentOS Isolate]
    AOS --> PI[Pi Agent Core]
    PI --> TOOLS[Shopify MCP Tools]
    PI --> LLM[LLM Provider]

Rivet manages the actor lifecycle (create, sleep, wake, terminate)
AgentOS provides the isolated execution environment for each actor
Pi Agent runs inside the AgentOS isolate

Performance

Metric	AgentOS	E2B	Daytona
Cold start	4.8 ms	440 ms	~1s
Memory	22 MB	1024 MB	~500 MB
Cost	1×	92×	254×

Why AgentOS for This Bot

Per-user isolation — each user’s agent runs in its own isolate
Fast cold starts — 4.8ms means instant scaling
Low memory — 22 MB per user means thousands per cluster
Integrated with Rivet — natural fit for actor-based architecture
Escalation — can mount full Linux sandbox if native tools needed later