Serverless

The Serverless service provides Function-as-a-Service backed by Firecracker microVMs. Each function invocation runs in its own lightweight VM for strong isolation. The Serverless Agent manages VM lifecycle, routing, and scaling.

Runtimes

## Features

• Sub-second cold starts — Firecracker boots microVMs in under 200ms
• Per-request scaling — each invocation gets its own isolated microVM
• Configurable resources — set memory (128 MB to 2 GB) and timeout (1s to 300s) per function
• S3-backed code — upload function code to AgentMetal Storage and reference it

Create a Function

agentmetal function create \
  --name process-image \
  --runtime python \
  --handler main.handler \
  --code s3://functions/process-image.zip \
  --memory 512 \
  --timeout 60

Invoke a Function

Via CLI

agentmetal function invoke process-image \
  --data '{"image": "photo.jpg", "width": 800}'

Via API

curl -X POST http://localhost:8080/v1/functions/process-image/invoke \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"image": "photo.jpg", "width": 800}'

Update Function Code

Deploy a new version of your function:

agentmetal function update process-image \
  --code s3://functions/process-image-v2.zip

View Logs

agentmetal function logs process-image --tail 50

List Functions

agentmetal function list

What the Agent Manages

The Serverless Agent performs these operations:

1. Firecracker VM lifecycle — creates and destroys microVMs for each function invocation, managing a pool of pre-warmed VMs for low-latency starts
2. Rootfs creation — builds minimal root filesystem images for each runtime containing the language runtime and your function code
3. Function routing — receives invocation requests and dispatches them to available microVMs, queuing requests when at capacity
4. Scaling — maintains a pool of warm microVMs based on invocation patterns, scales up during traffic spikes, and scales to zero during idle periods
5. Code deployment — downloads function code from S3, builds the rootfs, and makes it available for the next invocation
6. Monitoring — tracks invocation count, duration, error rate, and cold start frequency

Configuration

## Architecture

Each function invocation is fully isolated in a Firecracker microVM. MicroVMs boot in under 200ms and have their own kernel, network interface, and filesystem. This provides VM-level isolation with container-like performance.