Amp

Operator's take

Almost every tool in the agent-fleet lane shares one quiet assumption: the agents run on your machine, and the job is choreographing them — worktrees, ports, terminals, notification rings. Amp, Sourcegraph's coding agent, rejects the assumption instead of managing it. Its pitch line is the bluntest sentence in the lane: "Forget worktrees. Keep sending prompts instead."

The mechanism is Orbs: ephemeral remote machines where Amp agents run without supervision. Every new thread gets a fresh orb with your repository cloned, plugins and tools installed — "you set up nothing" — and the orb installs whatever the task turns out to need (browsers, databases, toolchains). When it's done, only the diff leaves the orb. Spawn unlimited ones at sizes from a $0.08/hour typo-fixer to a $2.13/hour 16-core box, billed by the minute; orbs pause themselves after five idle minutes and cost nothing paused. Agents spawn other agents in fresh orbs and hand work between them, and you can review diffs, inspect the agent's evidence, open a terminal into the same machine, and see E2E screenshots — from a browser, the CLI, or your phone, since nothing runs locally to tie you down.

The catch is what that freedom costs. Amp is closed and metered: the only tool in this lane where running the fleet is itself a line item, sized by the minute. "Forget worktrees" also means adopt Amp's cloud as where work happens — your repos are cloned into their machines, your mental model moves into their threads, and the local-first exit stories (a repo you fork, a runtime you run) don't apply. For teams that want the choreography gone badly enough to pay by the minute for it, that's the whole trade in one sentence.

What it's good at

• Zero-setup parallel machines — every thread gets a fresh orb with repo, plugins, and tools pre-installed; nothing runs on your machine and nothing needs configuring.
• Scales sideways without limits — unlimited orbs from tiny to 3xlarge, "no port conflicts, no disk space shortage, no memory pressure."
• Pays only while working — per-minute billing across the size range ($0.08–$2.13/hr), auto-pause after five idle minutes, paused orbs free.
• Agents that delegate — orbs aren't isolated boxes; agents spawn other agents in fresh orbs and exchange messages and files to hand off work.
• Review with the machine attached — browse files, review diffs, inspect the agent's evidence, open a terminal in the same orb, and get headless-browser E2E screenshots.
• Live preview URLs — Portals give every service in an orb a shareable URL (yourapp.onamp.dev).
• Wired into your stack — Slack @Amp mentions, webhooks from CI/GitHub/Linear, scheduled Automations that wake with full thread context, and Multiplayer for teammates in the same thread.

What it's not

• Not local — nothing runs on your machine, which is the pitch and also the dependency: your code lives in their orbs during the work.
• Not open source — a Sourcegraph product, not a forkable repo; the exit story is exporting your diffs and threads, not self-hosting the engine.
• Not flat-priced — the meter runs by the minute while orbs are active; a busy fleet is a real bill, unlike the free open-source locals in this lane.
• Not a worktree tool — it replaces the pattern rather than managing it; teams with heavy local tooling invested in the worktree model get less from it.