hckrnws

This looks great!

Looking to tinker with a few open source agentic workflows to integrate into my text editor so this will be helpful. Will def post feedback.

This is awesome, curious if you'd be open to an interview about your work on my youtube channel?

what youtube channel?

What are the requirements to run this?

Evidently not a lot- the RTX 3060 would be as low as I go for GOU horsepower with any sort of AI stuff.

At the other end of the spectrum... how noticeable a performance improvement would you get (if any) by running it on a Framework Desktop?

not sure if its due to the model or how it's setup, but say you tell it to find out about something, it goes searching on the web; I can see it crawl, but it's doing it stuck in an either forever loop or forever searching the web.

How is this secure? Can the agent run `rm -rf /` and destroy my machine by chance?

No it can't because we check the bash the AI try to execute against a list of pattern for dangerous command. Also all commands are executed within a folder specified in the configuration file, so that you can choose which files it has access to. However, we currently have no containerization meaning that code execution unlike bash could be harmful. I do think about improving the safety by running all code/commands within a docker and then having some kind of file transfer upon user validation once a task is done.

What if the agent were to create an alias to 'rm -rf' on my machine? I guess that would not have been blocked by your blacklist, right?

Well it can't use text editor, so it would have to use echo 'rm -rf' with a shell redirection to a file, which would be detected.

I guarantee you these controls are breakable the way you describe them.

Thats okay though! I realize this is a prototype/hobbyist solution which is unlikely to be attacked by a skilled adversary. Love the project!

If later on you want this to become safe for sensitive workloads you need to be way less confident. Just my 2¢.

I know, it's for local use, it's not hosted anywhere so the only adversary is yourself :)

I have not used this one yet but as a rule of thumb I always test this type of software in a VM, for example I have an Ubuntu Linux desktop running in Virtualbox on my mac to install and test stuff like this, which set up to be isolated and much less likely to have access to my primary Mac OS environment.

on linux make a non-root, limited shell login to start? avoid Windows

You can do a lot with seccomp filters that would stop even root messing things up too badly, down to path level io filtering unless I misremember

How does AgenticSeek handle agent communication and memory management?"

AgenticSeek only have agent communication for planning task but it isn't really communication. Planning task are simply a json list of task and how they related with each other. During execution output from each agents will be saved and "added" to the prompt of the following agents that depends on the information from the previous agent. Really dead simple but work quite well. For memory management, we store message as an array of {"role": "user", "content":...}. Where we take a more innovative approach if how we process message before adding to the history: We remove the <think>..reasoning..</think> pattern to keep only useful informations in the context (we don't really need the reasoning that was done on previous task for the current task). We also have memory summarization that is used when loading a previous session.