Project Roadmap and Positioning

[unbracketed]

Congratulations on your release @Ido-Levi. This is impressive work! It looks well thought out, and I hope the project serves you well.

I have a need for a very similar system, and just a few nights ago, before bed, I was taking down some notes about it. The next morning, I'm scrolling through Reddit and I see your release announcement. Incredible! I love your judicious use of diagrams throughout the documentation, very inspiring and it's a skill I'm trying to improve my game on significantly. In fact, the thing I'm building at the moment, and which I plan to use Hephaestus for, has a lot to do with treating diagrams and graphs as a first-class principle for organizing information and defining workflows. From what I can see, Hephaestus is going to give me a major boost in adding some of the core functionality that will tie it all together. I plan to use Hephaestus in two ways immediately:

1. To see how it feels being the primary driver of development for my project. I am a Claude Code power user with custom commands, skills, etc. And I think I will be able to adapt pretty well to what Hephaestus offers.
2. Integrating via the SDK into my project for creating and running Phases which can be used to provide actions and tasks in my project interface, and being able to hand off all of the agent orchestration.

I am curious how much thought you have put into, and/or if you are willing to share about the next steps for the project. More generally, how do you see it positioned in the overall landscape of agent-coordinating tools? For example, through one of the Reddit threads, I found out about the Archon project and it seems like there is a lot of surface-level overlap, with Hephaestus being more focused being an execution and coordination framework, while Archon provides more in terms of project task tracking and management but less control over the execution model. I also got the impression that Archon is more sophisticated with the RAG and memory features. No disrespect intended to either project if I'm way off base in my armchair assessment.

I agreed with another comment I saw that the project could benefit from some workspace or project management type features. Archon doesn't currently have any explicit SDKs though you can hack the API. I have my own workspace manager thing that was the beginnings of my own multi-agent coordinator from a few months ago that I could plug in as well. But if you already had some plan brewing, then I could hold off or coordinate with your work.

I'm also curious if there are architectural plans to make some components more pluggable like vector, storage, agent isolation, etc. Everyone has their own favorite choices of course; I suspect in time, a lot of pieces of the architecture can be adapted to support multiple backends, or options, or implementations. I have also implemented some basic tmux server and session controllers for coding agents, and think it's a great idea. I've been curious about going down the container route and using something like Dagger.io to dynamically create isolated containers - have you investigated anything like that?

Thanks so much for your hard work and for sharing it. I look forward to sharing a demo soon showing how I've integrated Hephaestus into my project.

[Ido-Levi]

Hey @unbracketed,
Thank you very much for trying out my project and giving your feedback!

About the next steps of the project - the project is still in its baby shoes, I've been working on this alone for about a month now (and on weekends and such) and it's been blowing my mind. It can be taken is so many different directions and I kind of want to try them all, but as far as "concrete" plans go, I definitely plan on building Hephaestus as a development tool by improving it's RAG and adding support for multi-workflow and multi-phase configurations, so Hephaestus could learn your repo as you use it more. And another thing that might interest me more is to see how it behaves for research.

Where Hephaestus differs from agentic frameworks like Archon is that the agents don't just move tickets across the board, but the phases architecture that let agents create previous tasks from previous phase, and that the tasks must follow the same phases (which are logical steps that are must in solving the goal, such as "plan -> implement -> validate") let agents at say the validation phase create tasks at previous steps if they need without working on it themselves is huge - because if a validation agent finds a new issue or has a new idea, he can just create the "plan" task and continue, and this would open an entire new branch - hence this is why it's semi-structured, because you don't have a predefined "do 1. 2. 3." but it's branching based on the goal and problem at hand, and this is SUPER strong at problems like doing bug bounties, where it can explore different combination of bugs that it finds (I had a test where it had over 20 agents running in parallel and it found quite a few bugs, and they've yielded a pretty nice bounty from HackerOne^).

And I definitely plan to have the agents run in isolated Docker containers and then run it with k8s, because right now the limitation of having it only run in my local machine means that running lots of agents in parallel makes my machine become a heater which isn't great - and I do want to get to the point where I can run 100s of agents in parallel - again, this is only a month of weekend work, and there is a lot more to come :)

I'm genuinely excited to see how you implement Hephaestus into your project, and if you have the time and you want to contribute to Hephaestus, I'll be more than happy to review and merge it!
As of right now, I've just published it and I'm waiting to see how people use it and get some feedback on what went right and what went wrong, and based on that adjust what I plan to work on.
As of right now, I'm running SWEBench-Verified on a new swebench workflow I've built, and it just started running but so far it has over 90% success rate on the first attempt, which is amazing! (still wayyy too early to celebrate, it has only ran only 15 issues^)