Remote Agent Workstation Architecture

Coding agents don't belong on your laptop.
They belong in Orbs.

An Orb is a dedicated remote machine where Amp agents run without supervision. Your repository, plugins, databases, and dependencies live in the cloud—so agents build, test, and migrate code across a parallel fleet while your laptop sits closed in your backpack.

Instant CLI execution: amp -ox "your prompt" --orb-size a1.small
Orb a1.medium
4 CPU • 8 GB RAM • 40 GB NVMe

The Laptop Trap: Local Execution

Running autonomous agents directly on your developer machine cripples your day:

  • Hardware Exhaustion: 6,000 RPM fan noise, battery drained from 100% to 15% in 40 minutes while Docker, Vite, and tests grind.
  • The Closed-Lid Catastrophe: Close your laptop to catch a train or step into a meeting, and the 45-minute migration halts and dies.
  • Serial Bottlenecks: Only 1 agent can safely run at a time without thrashing your working directory, git state, and port bindings.
THE ORB MODEL

The Amp Paradigm: Isolated Cloud Orbs

Agents work in disposable, dedicated cloud machines that you orchestrate effortlessly:

  • True Asynchronous Autonomy: Shut your laptop down. The orb finishes the migration, verifies with tests, and creates the PR remotely.
  • Parallel Fleet Scaling: Launch 5 orbs simultaneously. Five separate agents refactor five services in parallel for pennies.
  • 5-Minute Auto-Pause ($0/h): Sleep instantly when idle. Wake instantly upon access with 24-hour setup snapshot caching.

Four Ways to Launch an Orb

Trigger remote execution seamlessly from wherever you write software:

1. Web Thread

Click "Create New Thread" at ampcode.com, select a repository, and enter your prompt.

2. Terminal CLI

Run amp -ox "your prompt" or add --orb-size a1.small for explicit hardware sizing.

3. Interactive TUI

Open the command palette in the Amp TUI and choose thread: new in orb.

4. Plugin SDK

Invoke programmatically: agent.createThread({ executor: 'orb' }).

Architecture & Lifecycle

What actually lives inside an Orb?

Creating a new thread gives you a fresh Debian 12 virtual machine containing your code, your plugins, and the tools the agent needs. You retain full hands-on control while the agent executes remotely.

Repository Lifecycle Scripts

Deterministic configuration via repo scripts:

  • .agents/setup: Runs once while an orb is being prepared. Installs dependencies, generates files, seeds databases, verifies tools. Logs to /home/user/.cache/amp/logs/setup.log.
  • .agents/resume: Runs every time an existing orb wakes. Must be fast and idempotent, with a strict 10-second timeout. Logs to resume.log.
  • AGENTS.md: Repository instructions guiding agent decisions.

5-Min Auto-Pause & Snapshots

Economics built around dormant efficiency:

  • 5-Minute Auto-Pause: Idle orbs sleep automatically after 5 minutes of inactivity or when thread is archived. Sleeping orbs cost exactly $0.00/min.
  • Instant Resume: Any CLI access or portal request wakes the orb immediately.
  • 24-Hour Snapshots: After first setup, the sandbox state is snapshotted and reused for subsequent orbs, achieving sub-2-second spinups.

Full Hands-on Control & Sync

You are never locked out of the agent's environment:

  • Live In-Orb Terminal: Open an interactive bash terminal inside the remote orb anytime.
  • File Browser & PR Approval: Inspect, diff, and approve changes directly from the thread UI without pulling them down.
  • amp sync <thread-id>: Stream remote orb edits live to your laptop workstation in real-time while the agent keeps working remotely.
Debian 12 Toolchain

Interactive Pre-Installed Environment Explorer

Click any pre-installed package to inspect its role and test simulated CLI execution:

OS: Debian 12 (Bookworm) • Architecture: x86_64
user@orb-a1-medium:~
SSH Connected • Debian 12
user@orb:~$ agent-browser --version agent-browser v1.8.4 (Chromium 126.0.6478.126 headless) Ready for autonomous agent web navigation, form filling, and DOM extraction. user@orb:~$ which gh amp bun node psql redis-cli ffmpeg ripgrep /usr/local/bin/gh (Authenticated with repo scope) /usr/local/bin/amp (Amp executor daemon active) /usr/local/bin/bun (v1.1.20) /usr/bin/node (v20.15.0 LTS) /usr/bin/psql (PostgreSQL 16.3 ready on port 5432) /usr/bin/redis-cli (Redis 7.2.5 ready on port 6379) /usr/bin/ffmpeg (ffmpeg 6.1.1-3+b1) /usr/bin/ripgrep (rg 14.1.0)
Interactive Workday Simulator

The Remote Fleet Workstation

Don't just read about Orbs—orchestrate a live fleet. Spawn orbs, size them, assign tasks, and watch parallel agents complete a full workday while you sleep.

Select Workday Scenario:
INSPECTION RETICLE: X: 0.0, Y: 0.0
FOCUSED ORB: orb-84f2 (a1.large, active)
INTERACTION FRAME: fleet-canvas
STEER RETICLE: A / Left D / Right W / Up S / Down

Active Fleet Cluster (3 / 8 slots)

Click an orb card to inspect or trigger actions
Tasks Shipped 7 / 10
Engineering Time Saved 184 mins
Total Cloud Cost $0.42
Equivalent Dev ROI 99.4% Saved
First Remote Agent
Penny Pincher (Auto-Pause Saved > $1)
Swarm Commander (Parallel Execution)
Portal Pioneer (Wake via Web Request)
Lid Closed, Work Done
Head-to-Head Race

Beat The Laptop: Local Machine vs. 4-Orb Remote Fleet

Run the exact same heavy workload: "TypeScript 5.6 Migration + 3,800 Vitest Suites + Docker Postgres Re-seed". Experience what happens when you shut the laptop lid mid-flight.

Laptop Lid: OPEN

Local Developer Laptop (Single Agent)

Ready

Agent locks up local processes. Fans spin up to maximum throttle.

PROGRESS: 0% (0 / 3,800 tests)
CPU TEMP & FAN: 48°C • 0 RPM (Idle)
BATTERY CONSUMPTION: 100% (Normal)
ESTIMATED TIME: ~8m 45s

Amp 4-Orb Parallel Fleet

Ready

Work split across 4 remote a1.large orbs with gigabit interconnection.

PROGRESS: 0% (0 / 3,800 tests)
ORB CORES UTILIZED: 32 vCPUs across 4 Orbs
LAPTOP IMPACT: 0% CPU • 0 RPM • 0 battery drain
TOTAL RUN COST: $0.000
Hardware Sizes & Plans

Five Compute Sizing Tiers, Billed by the Minute

Pick the exact compute footprint needed for each prompt or lifecycle step. Orbs pause after 5 minutes of inactivity and resume in seconds.

Orb Size vCPUs RAM NVMe Disk Hourly Rate Per-Minute Rate Best Workloads
a1.tiny 1 Core 2 GB 40 GB $0.08 / hr ~$0.0013 / min Linting, documentation lookups, simple PR reviews
a1.small (Default) 2 Cores 4 GB 40 GB $0.17 / hr ~$0.0028 / min General feature coding, bugfixes, component tests
a1.medium 4 Cores 8 GB 40 GB $0.33 / hr ~$0.0055 / min Integration tests, Next.js / Vite compiles, Postgres seeding
a1.large 8 Cores 16 GB 40 GB $0.66 / hr ~$0.011 / min Large test suites, E2E browser runs, monorepo migrations
a1.xxlarge 16 Cores 32 GB 40 GB $1.32 / hr ~$0.022 / min Heavy C++/Rust compiles, massive data processing, batch ffmpeg

Agent Modes & Reasoning Tiers

Select the depth of autonomous reasoning per prompt:

Low Mode Fast

Lightweight edits, doc updates, and mechanical lint repairs. High speed, minimal reasoning latency.

Medium Mode Standard

Balanced reasoning for feature implementation, test drafting, and refactors across 3-8 files.

High Mode Deep

Deep architectural planning, multi-step migrations, and complex root-cause debugging across repos.

Ultra Mode Maximum

Autonomous multi-agent swarms, exhaustive boundary sweep testing, and zero-compromise monorepo upgrades.

Megawatt

For Individual Engineers

$20/ month

Everything needed for autonomous daily development.

  • 750 hours of a1.small remote orbs included
  • $20 of included agent reasoning usage
  • ✓ Access to Low and Medium agent modes
  • ✓ Web, Terminal TUI, and Mobile Phone access
  • ✓ 24-hour setup snapshot caching
Get Started with Megawatt
Unconstrained

For Teams & Enterprises

Pay-As-You-Go

Elastic scale with zero artificial concurrency caps.

  • ✓ Unlimited concurrency across all 5 orb tiers
  • ✓ Dedicated VPCs and self-hosted orb runners
  • ✓ Team audit logs, RBAC, and centralized billing
  • ✓ Custom Debian base images and private npm registries
  • ✓ 99.99% uptime SLA & dedicated support
Contact Enterprise Sales
Networking & Portals

Portals: Expose In-Orb Web Apps to Authenticated URLs

Portals expose HTTP servers running inside an orb as authenticated, secure URLs. Declare services in .amp/services.yaml, start them with amp orb services ensure, and test changes live in your browser.

Live Portal Request Simulation

When an orb is paused ($0/hr), visiting its authenticated Portal URL automatically wakes the machine, resumes billing, and serves the request in under 2 seconds!

Portal URL:
https://t-84f2-web.orb.ampcode.dev/dashboard
Orb Status: PAUSED ($0/h)
Response: Waiting for request...

.amp/services.yaml Config

# Declare long-lived services exposed via Portals
services:
  web:
    command: pnpm run dev --port $PORT
    port: 3000
    ready_endpoint: /api/health
  api:
    command: python -m uvicorn app:app --port $PORT
    port: 8000
  docs:
    command: bun run docs --port $PORT
    port: 4000

Each service automatically receives $PORT, $AMP_THREAD_ID, and $PUBLIC_URL. Services can reference each other's URLs seamlessly.