# Amp's users praise its cloud machines at a higher rate than any rival's

*Of the 1,323 people who posted about Amp in seven weeks, 131 praised orbs from their own use. They describe work that kept running through power cuts and trips, local setups they stopped maintaining, and rival cloud agents that fell short when they tried them.*

What users credit, Amp. By Feedback Bench, 15 October 2026. Feedback Bench 1.1. Posts 2026-08-31 to 2026-10-18; bench built 2026-10-19. Web: https://feedbackbench.com/coding-agents/#/articles/amps-users-praise-its-cloud-machines-at-a-higher-rate-than-any-rivals

On 10 September a developer in a power cut posted a short status report to Amp. "[i’m in the middle or a power outage… my laptop is at 3% and i’ll keep working normally, because @ampcode runs on my phone and i have orbs setup](https://twitter.com/1420340648117510149/status/2098098567168553083)," they wrote, and closed with "[what a time to be in tech](https://twitter.com/1420340648117510149/status/2098098567168553083)." The same user spent much of the seven weeks asking Amp for fixes, a graph of their runners and fewer incidents, which makes the praise easier to take at face value.

Orbs, the cloud machines Amp gives each thread, take the work off the user's own computer. The laptop can close, run flat or stay at home, and the threads carry on. It is the largest specific theme in Amp's praise that we counted. The next, threads that coordinate across machines, drew 42 people.

## How large it is, and where it stands

Feedback Bench read posts about 17 coding agents from 31 August to 18 October 2026. After we removed reply bots, Amp staff, vendors and people who hand out gift months or codes, 1,323 people had posted about Amp. Of those, 131 praised orbs from their own use, in 204 posts.

On the bench's own scoring, Amp and Devin lead on [cloud and remote sessions](https://feedbackbench.com/criteria/surfaces.cloud_sessions.md). Amp's score is twenty-one thousandths higher, still too little for the bench to separate the two. Amp is 1st of 7 on [mobile and remote access](https://feedbackbench.com/criteria/surfaces.remote_mobile.md), ahead of Claude Code and OpenCode. A simpler count, with one rule for every agent, shows a wider gap. Over the seven weeks, per 1,000 people who posted about each agent, 69 praised Amp's cloud sessions. Conductor came next with 33 and Devin with 17, so Amp's rate is 2.1 times the next. We did not read Claude Code's and OpenAI Codex's posts by hand; on the bench's own labels their rates are about four and two. That puts Amp first of the 17 agents on this measure.

**Exhibit 1: Amp draws the most cloud-sessions praise per person of any agent.** People who praised an agent's cloud or remote sessions, per 1,000 people who posted about that agent. One rule for every agent.

|  | Praise of cloud sessions per 1,000 people who posted about the agent |
|---|---|
| Amp | 68.6 (85 people) |
| Conductor | 33.1 (16 people) |
| Devin | 17.3 (77 people) |
| Cursor | 14.1 (166 people) |
| Factory | 10.4 (13 people) |
| Warp | 6.9 (2 people) |
| Cline | 2.5 (4 people) |
| Pi | 1.8 (7 people) |
| Zed | 1.4 (3 people) |
| OpenCode | 1.2 (18 people) |
| Kiro | 1.2 (1 person) |
| Google Antigravity | 1.1 (12 people) |
| GitHub Copilot | 0.9 (2 people) |
| Augment Code | 0.0 (0 people) |
| Grok Build | 0.0 (0 people) |

*31 Aug to 18 Oct 2026, all seven weeks of the window. Claude Code and OpenAI Codex are not shown; on the bench's labels their rates are about 4 and 2. Amp's own exclusions applied to Amp; Amp's posts are almost all X mentions of @AmpCode.*

Amp knows this about itself. Its homepage says "Send prompt, close laptop" and "Forget worktrees", and it carries a wall of user quotes on the same theme. So we read the 131 for what they add to that line: the moments, the machines they stopped needing, and the products they measured orbs against.

## The moments

Thirty-nine of the 131 describe working, or keeping work running, while away from their own computer. What struck us was how ordinary the settings are. One user spent "[a week in the yucatan using @ampcode on my iphone. no cell, wifi in one room. orbs were great.](https://twitter.com/610281461/status/2103902448830070922)" The same post goes on to list three annoyances with the phone app. Another shipped code at a conference "[on unreliable hotel conf wifi](https://twitter.com/784290/status/2098180599777280140)." A parent used the down time at a college move-in to clear some alerts. Orbs on a phone, they wrote, were "[still highly productive and allowed me to knock off a few sentry alerts](https://twitter.com/15023/status/2095867783326158947)."

The plainest version came from a user who tried another agent after a month on Amp. "[i forgot i had to keep my laptop open, closed it and all my threads died](https://twitter.com/287157700/status/2105983100924702740)," they wrote, and called the old way prehistoric.

Users credit two product decisions here. The first is that a thread runs in its own orb from the start, not on the user's machine, so closing the lid ends nothing. The second is that a thread started in the terminal continues in the iOS app without extra setup; Amp announced its iOS and macOS apps on 28 August, days before our window opened.

## The machine that got lighter

Twenty-five of the 131 say orbs replaced part of their local setup. They name the things they no longer handle: git worktrees, dependency installs, toolchains that drift. One wrote "[no more dealing with pnpm frying my mac or git worktrees!!](https://twitter.com/99500302/status/2106691478831440140)" Another, who had built a similar setup on their own machine with worktrees, found that "[worktree switch has a heavier context switch than orb + thread](https://twitter.com/2090734054928687104/status/2102117002349318514)" and could not quite say why.

A Rails developer had put off trying Amp because the project was hard to set up. Asked for a single bug fix, the agent set up the environment in an orb, saved what it learned for the next run, fixed the bug and opened a port for testing. "[my macbook is now just a microphone](https://twitter.com/19506770/status/2097323516689289251)," the developer concluded.

Nine people praised the cost, and their reason follows from the design. Amp bills orb time by the minute and pauses an orb five minutes after the agent stops or twenty minutes after the user's last interaction, whichever is later. A paused orb costs nothing, and the orbs docs add "even if it sleeps for weeks."

**Exhibit 2: What the 131 people who praise Amp's orbs credit.** People, not posts. One person can give more than one reason, so the rows add up to more than 131.

|  | People who praise orbs from their own use |
|---|---|
| Kept working away from their computer | 39 people |
| Orbs replaced worktrees or local setup | 25 people |
| Preferred orbs to a named rival cloud agent | 12 people |
| Preferred orbs to their own setup | 11 people |
| Said orbs cost little | 9 people |
| Praised orbs without one of these reasons | 64 people |

*31 Aug to 18 Oct 2026. Bots, staff, vendors and gift-month givers removed.*

## Measured against the alternatives

Twenty-three of the 131 compare orbs by name and prefer them. Eleven measure orbs against their own worktrees, terminal multiplexers or rented servers. Twelve measure them against another company's cloud agent: five against Cursor, four against OpenAI's Codex cloud, three against Claude Code's cloud sessions and one against a set of smaller sandbox services. One person named both Cursor and Codex, and one of the Codex posts credits Amp and Devin together.

The timing of these posts tells us something the homepage does not. The Cursor comparisons came mostly in the first days of September. The Claude Code ones came between 24 September and 1 October, and all four Codex ones came between 1 and 3 October, while users were trying what one called the new Codex cloud. The comparisons arrived in clusters, each while users were trying a rival's new cloud product.

After trying Cursor's cloud agents, one user found that "[everything about the cloud agents felt slower and less reliable then orbs](https://twitter.com/88027497/status/2095595925238640739)." Another tested Claude Code's cloud sessions "[just to figure out that in @ampcode orbs i never thought about cloud environments and that it takes 20 minutes just to install runtimes for my laravel project](https://twitter.com/1043795984856289280/status/2105747685097791808)." A third, on the new Codex cloud: "[better, still significantly worse than @ampcode](https://twitter.com/1978979128574644224/status/2105959826811429371)." Among rival cloud agents, the quotes on Amp's homepage name Cursor. None mentions Codex cloud or Claude Code's cloud sessions, and Amp's users have already made both comparisons.

## Who loves it most

Phone-first work and life away from the desk run through the largest group. People who had built their own parallel setup, with worktrees, multiplexers or a server, form the next. Many in both groups say they found orbs hard at first. One user spent 6 September "[having a hell of a time setting up a cloud agent](https://twitter.com/1751151/status/2096409568779567460)" and asked for a setup video. On 11 October the same person reported "[working off my phone all night with no hiccups and my mac mini completely offline](https://twitter.com/1751151/status/2109126187507204156)." The praise often comes after a few days of effort, and people say the effort was worth it.

## What to protect, and where it could reach

The users point to defaults more than to features. A fresh machine for every thread, an orb that stops costing money when it pauses, an environment the agent builds and remembers, and a thread that follows the user to the phone are the reasons people give. A change that made any of them optional would cut into the strength.

The same posts show where it could reach further. The phone is where users feel the wake-up time; for one user, checking a portal while on and off the phone meant a minute or two waiting for the orb to come back. Memory is the second edge, with several people outgrowing the smallest orb and asking to move work to a bigger one in one step. The third is Apple platforms. macOS cloud machines are among Amp's most frequent requests on the bench, and people building iOS apps still route builds to their own Macs. Users report that Mac orbs reached early access in October.

## Limits

Fifty-six people stated a problem with orbs in the same weeks, against the 131 who praised them, and eighteen people did both. The problems are outages, setups that failed, orbs that ran out of memory and the price of orb time; orbs draw on their own usage allowance, and some users ran it down. Praise outweighs complaint by about two to one, which is a good ratio but not a one-sided one. On the bench's scoring, Amp and Devin cannot be separated on cloud sessions.

Amp's posts are almost all X mentions of @AmpCode, where people address the company directly and praise runs high. We checked this two ways. Among X mentions alone, and as a share of each agent's praising authors, Amp still leads: 16 percent of the people who praise Amp praise its cloud sessions, against nine percent for Conductor and five percent for Devin.

## How we read this

We read every post about Amp in the window and checked each quoted author's other posts across all 17 agents. We removed reply bots, Amp staff, vendors, a rival's employee and people whose praise came with referral codes or gift months. A person counts when their own post praises orbs or Amp's cloud agents from their own use. Recommendations with no stated use, jokes, questions and wishes count on neither side, and the complaint side uses the same rule. We counted people, not posts. Large language models labelled the posts for the bench, and we read every counted post ourselves. The per-1,000 rate uses the bench labels for all seven weeks, with the same rule for every agent and Amp's own exclusions applied to Amp.

## Sources

- [Amp, homepage, read 2026-10-15 (last updated 2026-10-06)](https://ampcode.com/)
- [Amp, "Agents in Orbs", 2026-06-30](https://ampcode.com/news/agents-in-orbs)
- [Amp, "Amp on iOS & macOS", 2026-08-28](https://ampcode.com/news/amp-on-ios-and-macos)
- [Amp, Orbs documentation, read 2026-10-15 (last updated 2026-09-14)](https://ampcode.com/docs/orbs)
- [Amp, Pricing documentation, read 2026-10-15 (last updated 2026-09-14)](https://ampcode.com/docs/pricing)
- [Feedback Bench, Amp](https://feedbackbench.com/agents/amp.md)
- [Feedback Bench, Cloud and remote sandbox execution](https://feedbackbench.com/criteria/surfaces.cloud_sessions.md)
- [Feedback Bench, Mobile, remote-control and voice access](https://feedbackbench.com/criteria/surfaces.remote_mobile.md)
- [Feedback Bench, method](https://feedbackbench.com/#/method)