Github Agents!

​
This week, a colleague raised an important issue: he is running out of disk space due to multiple git worktrees.

The Problem is Git worktrees create a full copy of your repository for each worktree, which means:

∙ Each worktree recompiles all dependencies from scratch

∙ Storage usage multiplies quickly across multiple worktrees

∙ Local disk space becomes a serious constraint

A potential solution: Cloud Agents

It seems that even Anthropic is experimenting with this direction. I personally think this is how it will go, too.

We're already using cloud services for inference; local setups will never match the power of datacenters.

GitHub Agents offers an elegant alternative by moving the work off your local machine:

1. Creates a new branch in your GitHub repository

2. Spins up a fresh container via GitHub Actions

3. Runs the agent in that isolated container

4. Opens a pull request when the task is complete

It slots in very nicely with most developer workflows.

The Benefits are pretty clear:

∙ No local storage impact – Everything runs in the cloud

∙ Built-in sandboxing – Containers provide isolation by design

∙ Scalable – You can run as many instances as needed

But this comes with trade-offs.

It's yet another pile of config to manage.

Selecting a model is only available to Pro and Pro+ subscribers at the moment (apparently coming soon to Business and Enterprise tiers).

You're pretty out of the loop whilst things are running, although to me that is a potential benefit. And something my experiment will focus on, my theory is that giving excellent plans should help mitigate this.

You pay for the compute time of each running instance, but for our team, this cost is well worth avoiding the storage headaches.​​​​​​​​​​​​​​​​
​

It's how we already work

​
As Agents become even more capable, I think we will trust them more and inspect the results of their efforts rather than the process of them.

I’m seeing this in my own behaviour and my team's.

I manually test a feature before inspecting the diff. It’s essentially how we do code reviews anyway. My team sometimes uses ephemeral environments to do this, so they don’t even need to pull the code to their machines.

​

The experiment

Next week I’m going to lean into this and run an experiment. I’ll fully optimise this configuration, so GitHub Copilot agents can run independently in GitHub containers.

I’ll set up the ephemeral environments completely and optimise the context and tools available to the agents, so it's on par with my Claude code setup. Then I’ll see how much more productive I am, or if the underlying model is still not good enough.

It looks like I’m on a fast track to becoming a product manager.
​

I’ll let you know how it goes.

​
See you next Friday,

Dan

​

​ · ​