Agentenv Distributed Agent Environments

AgentENV Distributed Agent Environments

Skill by ara.so — AI Agent Skills collection.

AgentENV (AENV) is a distributed platform for running agent environments at scale using Firecracker microVMs. It provides fast snapshot/resume (<50ms boot, <100ms pause), native fork support, incremental snapshots to S3/distributed storage, and OCI image loading via overlaybd. Built in Rust, it powers agentic RL training workloads like Kimi K3.

Prerequisites

• Linux kernel 6.8+ (Ubuntu 24.04 recommended)
• /dev/kvm access for Firecracker
• Security Warning: AgentENV has no built-in authorization. Run only on trusted networks or behind an auth proxy.

Installation

Option 1: Install Script (Ubuntu 24.04)

Installs both server and CLI, starts server as systemd service:

curl -fsSL https://raw.githubusercontent.com/kvcache-ai/AgentENV/main/scripts/install.sh | sudo bash
sudo systemctl start aenv

Check status:

sudo systemctl status aenv

Option 2: Docker

curl -fsSL https://raw.githubusercontent.com/kvcache-ai/AgentENV/main/scripts/docker-setup.sh | sudo bash
docker pull ghcr.io/kvcache-ai/aenv-server:latest
docker run -d --privileged -v /dev:/dev -p 8000:8000 ghcr.io/kvcache-ai/aenv-server:latest

CLI Only (Linux/macOS, x86_64/arm64)

If server is on another machine or using Docker:

curl -fsSL https://raw.githubusercontent.com/kvcache-ai/AgentENV/main/scripts/install-cli.sh | bash

Authentication

Configure CLI to point at your server:

aenv auth
# AENV server URL [http://localhost:8000]: http://127.0.0.1:8000
# API key: dummy

For production, set:

export AENV_SERVER_URL=http://your-server:8000
export AENV_API_KEY="${AENV_API_KEY}"

Core Concepts

Templates

Templates are OCI-compatible images converted to AgentENV format. They serve as base images for sandboxes.

Sandboxes

Sandboxes are running microVM environments created from templates. They can be paused, resumed, snapshotted, and forked.

CLI Commands

Template Management

# Pull Docker image as template
aenv pull docker.io/library/ubuntu:22.04 --name ubuntu
aenv pull python:3.11-slim --name python311

# List templates
aenv template list
aenv template ls  # alias

Sandbox Lifecycle

# Start interactive sandbox (attaches shell)
aenv start ubuntu

# Start detached (returns sandbox ID)
aenv start ubuntu --detach
# Output: sandbox-abc123def456

# List running sandboxes
aenv ls
aenv list --output json  # JSON output for scripting

Sandbox Operations

# Attach to running sandbox
aenv cn sandbox-abc123def456

# Execute command in sandbox
aenv exec sandbox-abc123def456 ls -la /
aenv exec sandbox-abc123def456 python3 script.py

# Pause sandbox (free CPU/memory)
aenv pause sandbox-abc123def456

# Resume paused sandbox
aenv resume sandbox-abc123def456

# Extend TTL (timeout in seconds)
aenv timeout sandbox-abc123def456 600  # 10 minutes from now
aenv timeout sandbox-abc123def456 3600  # 1 hour

# Delete sandbox
aenv delete sandbox-abc123def456
aenv rm sandbox-abc123def456  # alias

HTTP API Usage

AgentENV exposes a REST API compatible with E2B. Base URL defaults to http://localhost:8000.

Create Sandbox

curl -X POST http://localhost:8000/sandboxes \
  -H "Content-Type: application/json" \
  -H "X-API-Key: ${AENV_API_KEY}" \
  -d '{
    "template": "ubuntu",
    "timeout": 600
  }'

Response:

{
  "sandbox_id": "sandbox-abc123def456",
  "status": "running"
}

Execute Command

curl -X POST http://localhost:8000/sandboxes/sandbox-abc123def456/exec \
  -H "Content-Type: application/json" \
  -H "X-API-Key: ${AENV_API_KEY}" \
  -d '{
    "cmd": ["python3", "-c", "print(\"Hello from AgentENV\")"]
  }'

Response:

{
  "exit_code": 0,
  "stdout": "Hello from AgentENV\n",
  "stderr": ""
}

Pause/Resume

# Pause
curl -X POST http://localhost:8000/sandboxes/sandbox-abc123def456/pause \
  -H "X-API-Key: ${AENV_API_KEY}"

# Resume
curl -X POST http://localhost:8000/sandboxes/sandbox-abc123def456/resume \
  -H "X-API-Key: ${AENV_API_KEY}"

Delete Sandbox

curl -X DELETE http://localhost:8000/sandboxes/sandbox-abc123def456 \
  -H "X-API-Key: ${AENV_API_KEY}"

E2B Compatibility

AgentENV implements the E2B API. Use the official E2B SDK without code changes.

Python SDK

pip install e2b

import os
from e2b import Sandbox

# Point to AgentENV server
os.environ["E2B_API_URL"] = "http://localhost:8000"
os.environ["E2B_API_KEY"] = os.getenv("AENV_API_KEY", "dummy")

# Create sandbox from template
sandbox = Sandbox(template="ubuntu", timeout=600)

try:
    # Execute commands
    result = sandbox.commands.run("ls -la /")
    print(result.stdout)
    
    # Write and execute Python script
    sandbox.filesystem.write("/tmp/test.py", "print('Hello from E2B on AgentENV')")
    output = sandbox.commands.run("python3 /tmp/test.py")
    print(output.stdout)
    
finally:
    sandbox.close()

TypeScript SDK

npm install @e2b/sdk

import { Sandbox } from '@e2b/sdk';

process.env.E2B_API_URL = 'http://localhost:8000';
process.env.E2B_API_KEY = process.env.AENV_API_KEY || 'dummy';

const sandbox = await Sandbox.create({ 
  template: 'ubuntu',
  timeout: 600 
});

try {
  const result = await sandbox.commands.run('ls -la /');
  console.log(result.stdout);
  
  await sandbox.filesystem.write('/tmp/test.js', 'console.log("Hello from E2B on AgentENV")');
  const output = await sandbox.commands.run('node /tmp/test.js');
  console.log(output.stdout);
} finally {
  await sandbox.close();
}

Snapshot and Fork Patterns

Creating Checkpoints

Snapshots complete in <100ms even with heavy disk modifications:

# Via CLI (requires API call or direct server access)
curl -X POST http://localhost:8000/sandboxes/sandbox-abc123def456/snapshot \
  -H "X-API-Key: ${AENV_API_KEY}" \
  -d '{"name": "checkpoint-training-epoch-5"}'

Forking Environments

Fork a running sandbox for parallel workflows:

curl -X POST http://localhost:8000/sandboxes/sandbox-abc123def456/fork \
  -H "X-API-Key: ${AENV_API_KEY}"

Response:

{
  "sandbox_id": "sandbox-xyz789ghi012",
  "parent_id": "sandbox-abc123def456"
}

Agent Training Workflow Example

import os
import time
from e2b import Sandbox

os.environ["E2B_API_URL"] = "http://localhost:8000"
os.environ["E2B_API_KEY"] = os.getenv("AENV_API_KEY")

def train_agent_episode(template: str, agent_code: str, episode_num: int):
    """Run single training episode in isolated sandbox."""
    sandbox = Sandbox(template=template, timeout=3600)
    
    try:
        # Install dependencies
        sandbox.commands.run("pip install numpy gymnasium torch")
        
        # Deploy agent code
        sandbox.filesystem.write("/workspace/agent.py", agent_code)
        
        # Run training episode
        result = sandbox.commands.run(
            f"python /workspace/agent.py --episode {episode_num}",
            timeout=1800
        )
        
        # Collect metrics
        metrics = sandbox.filesystem.read("/workspace/metrics.json")
        
        return {
            "episode": episode_num,
            "exit_code": result.exit_code,
            "metrics": metrics,
            "sandbox_id": sandbox.id
        }
        
    finally:
        sandbox.close()

# Run parallel episodes
agent_code = open("my_agent.py").read()
results = []

for i in range(10):
    result = train_agent_episode("python311", agent_code, i)
    results.append(result)
    print(f"Episode {i} completed: {result['metrics']}")

Distributed Cluster Deployment

Docker Compose

Create docker-compose.yml:

version: '3.8'

services:
  aenv-server:
    image: ghcr.io/kvcache-ai/aenv-server:latest
    privileged: true
    volumes:
      - /dev:/dev
    ports:
      - "8000:8000"
    environment:
      - AENV_BIND_ADDRESS=0.0.0.0:8000
      - AENV_STORAGE_BACKEND=s3
      - AWS_ACCESS_KEY_ID=${AWS_ACCESS_KEY_ID}
      - AWS_SECRET_ACCESS_KEY=${AWS_SECRET_ACCESS_KEY}
      - AWS_REGION=${AWS_REGION}
      - AENV_S3_BUCKET=${AENV_S3_BUCKET}
    restart: unless-stopped

Deploy:

docker-compose up -d

Kubernetes (Helm)

helm repo add aenv https://kvcache-ai.github.io/AgentENV/charts
helm install aenv aenv/agentenv \
  --set storage.backend=s3 \
  --set storage.s3.bucket="${AENV_S3_BUCKET}" \
  --set storage.s3.region="${AWS_REGION}"

Configuration

Environment Variables

# Server configuration
export AENV_BIND_ADDRESS=0.0.0.0:8000
export AENV_LOG_LEVEL=info  # debug, info, warn, error

# Storage backend (local, s3, distributed-fs)
export AENV_STORAGE_BACKEND=s3
export AWS_ACCESS_KEY_ID="${AWS_ACCESS_KEY_ID}"
export AWS_SECRET_ACCESS_KEY="${AWS_SECRET_ACCESS_KEY}"
export AWS_REGION=us-west-2
export AENV_S3_BUCKET=my-aenv-snapshots

# Resource limits
export AENV_MAX_SANDBOXES=100
export AENV_DEFAULT_TIMEOUT=600  # seconds
export AENV_MAX_MEMORY_MB=2048
export AENV_MAX_VCPUS=2

# Overlaybd cache
export AENV_CACHE_SIZE_GB=50
export AENV_CACHE_DIR=/var/cache/aenv

Advanced Patterns

Custom Template Creation

# Pull base image
aenv pull ubuntu:22.04 --name base-ubuntu

# Start sandbox and customize
SANDBOX_ID=$(aenv start base-ubuntu --detach)

# Install software
aenv exec $SANDBOX_ID apt-get update
aenv exec $SANDBOX_ID apt-get install -y python3-pip git

# Create template from running sandbox
curl -X POST http://localhost:8000/templates \
  -H "X-API-Key: ${AENV_API_KEY}" \
  -d "{\"name\": \"custom-python\", \"sandbox_id\": \"$SANDBOX_ID\"}"

# Clean up
aenv rm $SANDBOX_ID

Long-Running Agent with Auto-Pause

import time
from e2b import Sandbox

sandbox = Sandbox(template="ubuntu", timeout=86400)  # 24h

try:
    while True:
        # Do work
        result = sandbox.commands.run("python agent_step.py")
        
        # Pause during idle periods
        if result.stdout.strip() == "idle":
            sandbox.pause()
            time.sleep(60)
            sandbox.resume()
        
        time.sleep(5)
        
finally:
    sandbox.close()

Batch Processing with Sandbox Pool

from concurrent.futures import ThreadPoolExecutor
from e2b import Sandbox

def process_task(task_id: int, template: str):
    sandbox = Sandbox(template=template, timeout=600)
    try:
        result = sandbox.commands.run(f"python process.py --task {task_id}")
        return {"task_id": task_id, "output": result.stdout}
    finally:
        sandbox.close()

# Process 50 tasks across 10 parallel sandboxes
with ThreadPoolExecutor(max_workers=10) as executor:
    futures = [executor.submit(process_task, i, "python311") for i in range(50)]
    results = [f.result() for f in futures]

print(f"Processed {len(results)} tasks")

Troubleshooting

Sandbox Won't Start

Check KVM access:

ls -l /dev/kvm
# Should show: crw-rw---- 1 root kvm

# Add user to kvm group
sudo usermod -aG kvm $USER

Check kernel version:

uname -r
# Should be 6.8.0 or higher

Server Not Responding

Check service status:

sudo systemctl status aenv
sudo journalctl -u aenv -n 50

For Docker:

docker ps
docker logs <container-id>

Template Pull Fails

Check overlaybd setup:

# Verify overlaybd is running
sudo systemctl status overlaybd

# Check disk space
df -h /var/cache/aenv

Sandbox Timeout Issues

Extend timeout before starting:

# CLI
aenv start ubuntu --detach
aenv timeout sandbox-abc123def456 7200  # 2 hours

# API
curl -X POST http://localhost:8000/sandboxes \
  -H "X-API-Key: ${AENV_API_KEY}" \
  -d '{"template": "ubuntu", "timeout": 7200}'

Memory/CPU Limits

Configure per-sandbox resources:

curl -X POST http://localhost:8000/sandboxes \
  -H "X-API-Key: ${AENV_API_KEY}" \
  -d '{
    "template": "ubuntu",
    "mem_size_mib": 4096,
    "vcpu_count": 4
  }'

Check Sandbox Logs

# Via CLI
aenv exec sandbox-abc123def456 cat /var/log/syslog

# Via API
curl http://localhost:8000/sandboxes/sandbox-abc123def456/logs \
  -H "X-API-Key: ${AENV_API_KEY}"

Performance Optimization

Pre-warm Templates

Pull templates before workload starts:

aenv pull ubuntu:22.04 --name ubuntu
aenv pull python:3.11-slim --name python311
aenv pull nvidia/cuda:12.2.0-runtime-ubuntu22.04 --name cuda

Use Pause/Resume for Idle Periods

Pausing releases CPU and most memory in <100ms:

# Pause during long waits
sandbox.pause()
time.sleep(300)  # Wait for external event
sandbox.resume()

Leverage Forking for Parallel Workflows

Fork instead of creating new sandboxes for faster startup:

# Create base sandbox with loaded dependencies
BASE_ID=$(aenv start python311 --detach)
aenv exec $BASE_ID pip install torch numpy pandas

# Fork for parallel tasks
FORK1=$(curl -X POST http://localhost:8000/sandboxes/$BASE_ID/fork | jq -r .sandbox_id)
FORK2=$(curl -X POST http://localhost:8000/sandboxes/$BASE_ID/fork | jq -r .sandbox_id)

Resources

• Documentation: https://kvcache-ai.github.io/AgentENV/
• GitHub: https://github.com/kvcache-ai/AgentENV
• Issues: https://github.com/kvcache-ai/AgentENV/issues
• E2B SDK: https://github.com/e2b-dev/e2b