I built my own software factory, and I open-sourced it. It's called Sandcastle. Here's how to use it:

Here's the repo:

And if you want to learn how to do real engineering with coding agents, head to my newsletter:

I truly want to understand. What makes this better than an Automaker, Aperant or Paperclip? So do you work with your main agent to run this workflow? I'm going to try this out for sure, but parallel agents isn't super difficult anymore. Is this meant to work with customizable coding processes?

Give me the rundown on those tools, and I'll tell you how Sandcastle is different. Essentially, Sandcastle is just a primitive - a run() function that lets you build your own workflows.

Ok custom workflows with agents. So my workflow is — read project -> get next task from linear -> open merge request -> plan task -> do subtasks -> multiple commits sub tasks to merge request -> done? -> handoff to QA -> qa review -> qa feedback -> agent fix -> qa confirm -> then updates merge request to ci ready -> ci agent reviews -> if green, then merge -> update gitlab Do all this with my discord agent on a cron. It just runs the workflow. It could be just like that? Everything via cli agents?

Yep, exactly

Thanks @mattpocockuk - here's some feedback from gpt 5.5 xhigh -

The planning + worktree/branch strategy for parallel agents is a nice touch. But running in Docker/Podman with bind mounts (no network controls, no read-only rootfs, etc.) is containerized execution, not actual isolated sandboxing, so it can mislead folks.

Yep, it's on you to add layers on top of that. But it's pretty simple in the Dockerfile

Love this! I built @afkode_ai which is sharing the exact same philosophy. this is a feature running completely AFK for 105 tasks. Each task also starts with a fresh curated context so it gets only what's needed Great Job!

When you build your own factory, you own the speed and quality. Smart move open-sourcing that blueprint.

How would you say that this is different than Symphony Matt?

I would love to see you samples of completed projects, with details on what prompts were used

Incredibly well thought out, thank you for sharing and letting us use it!

bellissimo! peccato per github...

Hey, I just came to search for you. I'm what you guys would call a normie. I'm a full time carer so just trying to automate some things. I came across your git and Claude/Hermes saying it's the best thing have seen and helps a lot with my time/work. Many thanks to you man

Appreciate it pal! Hope you get some time off, full-time caring is no joke.

Dont ever underestimate the value you and others bring, im prepared to read/learn & graft, but the time you have saved me & utilisation I now have is immense. I care for wife & son this allows me to hit a deadline I have to finish a book for my 1st possible deal-beyond grateful

Matt the gift that keeps giving 🥹

Can i use sandcastle with ?

How do u manage this on codebases with many different engineers? Do you jsut modify the label to be a certain user so my agents will only look at the issues w/ the label for me?

Correct, or you collaborate on running the same agent.

Built my own software factory last October. Vibe-coded the whole thing in a weekend. By week 3 it had shipped 4 products, deprecated 2 of them, and filed 3 trademarks on names I hadn't approved. Two of the trademarks were for competitors. Legal is aware.

i dont even care if it works, the pitch is great! we need more stuff like this , not tweak here tweak there, but "a product" which reasonable respected vibedevs can use to become slightly less degenerate

very cool! I built a similar library but a bit more focused on workflow primitives vs. agents + sandboxing, but I see a lot of similar patterns. awesome to see folks converging on the same ideas :)

This is dope and looks really clean. How long did it take you to build?

I've been doing something similar, but running it on GH self-hosted runners.

Wow very nice stuff, would be fun to pair it with autoresearch/autodiscovery framework, then it can ship real products, not just software but an entire businesses.

Do you use basic vs code ide?

@makosamunyaa

What’s the process of coming up with the issues/tasks for the agent?

In my skills, it's: /grill-with-docs, then /to-prd, then /to-issues

Thanks! Going to test this flow out on a project I’ve been wanting to build for quite some time

very nice, might have matching parts

Does this work with codex or is it exclusive to claude code?

@tricksahoy1020 It's agent agnostic, Codex and CC and Pi are supported.

Like... Sandcastles in the sand?

You should check out Fusion.

I’ve got my own Docker-based AI sandbox setup that mostly works, but it’s been a constant source of pain / bugs, so I’d love to replace it with Sandcastle if I can. My normal workflow is one primary repo, but I often mount a couple of extra repos for context e.g. my Obsidian vault for daily notes, or another codebase as reference. I don’t usually need Sandcastle to manage branches/commits/deps for those extra repos. I mostly just need them visible to an interactive pi session, and ideally for repo-local AI config/instructions to be discoverable when the agent works in those dirs. Is that a reasonable Sandcastle use case via Docker mounts, or am I stretching it beyond the intended model?

maaan. this is good. Do you use skills in there or only the .md filed for prompts? Looks elaborate on env management but light on agents/skills/plugins

You can just mount your skills directory over to the container, simple

Did you try modal sandboxes? Look forward to comparing this against it

Crazy cool thanks for sharing

The question I have is how much does this cost?

Been looking for something exactly like this - every tool I've tried for automation still needs constant babysitting. How's the maintenance overhead been for you?

wait so what's the actual differentiator here vs something like nix flakes or just docker compose for the reproducibility angle

Those are sandbox mechanisms, Sandcastle handles running the agent inside them, tracking events, token usage, resumability, and worktree management on top.