Remote Agents

Written by Kyle Satti

~8 min read

Posted 5 days ago

This weekend I used GitHub Copilot's "coding agents", by which they mean agents that run on github.com, not the "agent mode" in VS Code. They're a ton of fun! I thought I'd share a little bit about them and how they compare to similar offerings from Cursor, Anthropic and GitHub to share with you.

If you want to skip right ahead to Copilot coding agents in action, you can checkout a few GitHub Issues here that I assigned GitHub Copilot to handle:

1. CI setup for my ESP32 project
2. Add automatic reconnection to Wi-Fi and WebSocket
3. Add more requests to an API client

Remote Agents are AI agents that run on a remote server as opposed to the usual model of today where your coding agent runs in your terminal (e.g. Claude Code) or IDE (e.g. Cursor).

Imagine you're talking to ChatGPT and you can ask it to do something for you, like fix a bug a implement a new feature, but instead of just giving you the code, it actually goes away to its own remote container, researches and plans your prompt and then creates a PR for you to review when you're ready.

For example, this remote approach means that you can spin up an agent from your iPhone and have a PR ready for you to review when you get back to your desk.

They're awesome for getting your ideas out while you're on the go and having AI do some initial research and implementation for you in the background and they're also great for contained agent sessions that don't require your machine to be awake/powered.

Below are a few screenshots of me creating two different GitHub Issues (1, 2) and assigning them to GitHub Copilot Coding Agents to handle, followed up by notifications that Copilot has created the pull requests for me to review 10-20 minutes later.

What I love about how it works is that, like Claude Code, it details a todo list of steps its going to take to get your requet resolved.

Also like Claude Code, you can give the remote agent guidelines for your repo:

• How to build and test your code
• Guide its workflow on each remote session (e.g. I like my remote agent runs to also commit a build log file of what it did, so the next agent has more context)
• Style guides etc. etc.

I've found remote agents to be an awesome concept for a couple of reasons:

1. anywhere, anytime – prompt and go. Start on mac, check in on iPhone. Or prompt when you have an idea on your iPhone while you're on the go and then open your mac when you get home to see what it did
2. isolated & safe – because the agent is off running in some container, it can go ham on a branch of its own and not break my machine, and it can run tools without me approving each of them basically zero risk.

For more on getting to know Remote Agents I recommend this section of a @wesbos YouTube video where he talks about their introduction and utility.

1. They cost quite a bit more than local agents (see below, TL; DR; ~$2-$5 per "reasonable sized" PR)
2. For more complex tasks, you might want to have quicker/closer access to interupt and guide the iteration. Copilot's UX for this today doesn't make that so easy.
3. Sub-agents don't seem to be supported yet, and even if they were, the cost would probably be pretty crazy

Cursor and Claude also have offerings for these types of agents, both calling them "Background Agents".

The terms "Background Agents" and "Coding Agents" could be misinterpreted as something you're running on your local machine (e.g. with Claude Code in the terminal or in the Cursor IDE), so I prefer to call them "Remote Agents", since that makes it clearer that they're not running on your machine.

GitHub Copilot, Cursor and Claude Code all offer Remote Agents, but only Copilot offers them within their subscription (perhaps not for long though... 👀 read on)

GitHub "Coding Agents" use GitHub's "premium requests". Each subscription tier has allowances for premium requests and you can optionally enable pay-per-request after your allowance is used up.

Here's an example PR that GitHub Copilot Agents made to a ESP32 repo of mine that implemented automatic wi-fi / websocket reconnection. It made 48 premium requests in the process of doing this. At the time, I'd already used up my allowance on the Pro tier, so this translated into a cost of $2.

Here's the allowance of each tier:

• Free: 50 premium requests
• Pro: 300 premium requests
• Pro+: 1500 premium requests

If I paid for the Pro+ tier, GitHub Coding Agent could make me roughly 31 of the PR I linked to above per month for $39. This would use up all 1500 of my monthly requests, though, leaving me with none for other features that require them such as Copilot Chat, Agent mode in Copilot Chat, Copilot Code review, Copilot Extensions and Copilot Spaces (read more here).

Cursor introduced background agents in mid-May 2025 but sadly their usage is not included in any of their plans.

Cursor plans merely give you "access" to Background Agents. Usage of Background Agents is charged at "API pricing" (i.e usage-based pricing). It seems most of us seem to prefer a monthly flat rate rather than paying per usage since anxiety of innocently blowing all your credits on a prompt that gets out of hand is a very real thing. Hopefully Cursor reconsider this.

A mere 7 days after Cursor announced their Background Agents, Anthropic too added "Background Agents" via an official GitHub Action.

Sadly, like Cursor, Claude Code does not currently included usage of this GitHub Action through any of its subscriptions. The GH Action uses Claude's API, for which you have to buy credits... booooo usage-based pricing.

HOWEVER! Some clever people (see this GitHub issue on the official GH action repo where folks are asking Anthropic for access the GH Action via their plans instead of via dev credits) have got together and forked the official Claude GitHub Action to work with the Claude Max Subscription!

This fork states compatibility with the Max plan, but I tried it out with my Claude Pro plan anyway; sadly this resulted in an error about an invalid model name being passed to Claude.

Before blindly upgrading to Claude Max, I started to dig deeper into whether this error was indeed caused by me not having a Max plan, which is when I noticed this comment in an issue on the GH action fork where the author shared that Anthropic will soon announce that their Pro and Max plan will work with this Remote Agent GH Action 🎉

If you want to keep things local and don't mind keeping a machine of yours online to run agents, you can use Vibe Tunnel.

Even if you don't use Vibe Tunnel, I recommend following their developers @badlogic, @mitsuhiko and @steipete who are sharing a lot of amazing tips and details of getting the most out of AI agents.

Remote Agents are something I'm really excited about for parallelizing work without it being bound to my machine and prompting from my iPhone on the run.

That said, it's obviously not going to be a replacement for local agent development when you want more granular control over the iteration or to support other workflows such as sub-agents.

I'll personally be waiting for Anthropic to add Remote Agent functionality to their subscriptions and will be happy to upgrade from the Pro to the Max plan to get it.