I’ll send this to the team at my company. No one likes feeling vendor locked in on artifacts but they are great at being like “hey I had an agent cook on this topic over night, here are the results, and what we need to make a decision on”.

Just some feedback for corporate use - it’s not clear to me from the landing page if it’s really intended for internal company use and if it’s hard tied to CloudFlare. I.e. if there was a “here’s how you can deploy it on AWS in 2 minutes” option I would just toss it on our dev env and show people a POC.

With open source projects we’re increasingly worried about supply chain type risks - it would be good if there was some boilerplate on “this is how you know this can be deployed internally and have zero risk of phoning home etc”

Very nice!!! I actually built something much more primitive as my workflow too. I'll definitely send the robots to learn from you and what you did!

(Edited after research: I picked up your ideas of search and versioning. Very nice! Thank you!)

interesting

seems like they "gadgets" concept of Cloudflare OS [1] but decoupled front Cloudflare OS chat interface.

It even runs on Durable Object like CFOS

1. https://os.cloudflare.app/

It's a skill like anything. Context switching on this level is initially exhausting, but do it enough and it becomes natural.

I've found an agent to manage agents, plus aggressively creating skills/workflows helps.

The main agent is the entry point:

"review this PR" should open a new workspace and spin up one or more agents to do whatever review workflow you like.

"Implement ticket xyz" should open a new space, read the ticket content, plan/write/test/review/create a PR without much managerial input. You'll still need to answer questions and check work between steps, but all of the busy work can be done by your agent(s) with skills. If they get it wrong, update the skill.

Same for "research this bug report/product question/error trace".

I guess the main benefit of herdr is it essentially flags the state of agents and what needs attention. So you can have multiple ones running and quickly go to the ones that need an input.

I think there _is_ a way to do this nicely in a GUI btw, which I've been toying with as a side project. Basically... build it as a trello style board.

And it's not 12 projects, its multiple tasks, e.g. 'look at this crash for me', 'fix this extension in my pi env that is doing something bad', 'debug this high memory usage', etc. etc. It's easy to just fire tasks off and come back to them later. The project ones usually require more human focus ofc.

Now if only I could automate meetings like this.

From the people around me doing this - they aren't keeping up, they aren't looking at the finished project/product, then throwing it over the wall for others to triage. Rinse and repeat. Definitely fits some people's personalities, but not mine.

Cc desktop supports folders, named tabs, archived sessions, subagents with two way comms between them and their spawner.

You cannot put them next to code being changed but the code viewers are contextual to the session.

I am struggling to see a major difference let alone major advantage of herdr