Comment by joshstrange | Hacker News Reader

Comment by joshstrange | Hacker News Reader

Parent Full thread

Ok, I've been running it for a couple hours and below are my thoughts. Please note that I do find it fascinating even if most of what I'm about to say is complaining about the parts I like less.

- Pipelines, not agents. Roles, not agents. I really don't like the idea of defined agents with their own prompt. I want to define roles and spin up N agents with that role. Furthermore I want pipelines "Plan -> Review Plan -> Approval Gate -> Develop -> Code Review + Fix loop -> QA -> Approval Gate -> Merge -> [Ship]". I don't like the work just bouncing around seemingly randomly

- Settings don't seem to save/persist? Or some of them don't. I couldn't let "Michael" spin up agents on "his" own and then randomly he did it even though the setting was still off. Settings has the normal LLM jank I've seen.

- macOS Notifications are broken, they send for any little reason, and then they don't send when you're actually needed. It's like each agent finishing a round causes a notification.

- Speaking of missing notifications, the _most important_ screen to me is the "Ask Me" tab under "Michael", where they ask questions (more on that later) but there is zero indication that anything is waiting for you. You have to dig into it yourself.

- The "Ask Me" tab is great.... when it works. I've had to unstick agents or answer questions they were waiting on answers for

- Trying to be too cute, it was cute for a minute, now I don't care (and I _love_ The Office). I want a more utilitarian view. I want to see questions, plans, be able to inject new ideas, and a small overview of what each agent is doing. I don't need half the screen taken up with a "game ui".

- Why no clear? I don't understand at all the idea of them keeping context. Maybe I'm missing something and I shouldn't be using persistent agents except for more persistent jobs (like Michael's?).

It's an interesting concept, very "Gas Town", and it make me want to write my own that does more of what I'm looking for but I don't have the time (or tokens) currently to take on another project. My current best approach of herdr+6-10 Claude Code sessions feels like it works better than this and keeps me close enough to the decisions I want to make.

I think you and I are cut from the same cloth. Role-based pipelines and coordination with scoped permissions generated at runtime and the ability to communicate is the end game in my opinion, for now at least. Orchestration finally clicked with me when I tried out Gas Town six months ago and I immediately created a system to offload it to the cloud by using a combination of GCP CloudRun + Firestore (for temp file storage and to fill role of beads). Right when I finished ironing out the kinks Anthropic laid down the hammer on how users can use subscriptions and I abandoned it.

Now I am using custom pipeline/messaging control plane that spawns microvms with scoped creds minted at runtime. I had to make my own control plane specifically because everyone is focused on agents with personalities rather than roles and pods of workers. I like everything about Buzz except for that.

I'm slowly inching towards my own control plane because my skill-based workflows (not Claude Code Workflows) are brittle but still work better than anything else I've tried. Enforcing the pipeline has been mostly successful but I'd greatly prefer something more structured.

I've been dipping my toes into the waters of higher-level orchestration for a while now but keep scurrying back to driving CC mostly directly. Any extra layers I've added have just caused too much waste (time & tokens) or otherwise produced inconsistent results. I want to "live" in the decision making space. I want to be consulted for decisions (with context provided so I can context-switch easier), I want to review plans, I want to review the actual UI (when there is any), and I want to be able to easily inject new ideas/tickets into the machine (and probably so way to adjust issue priority).

Thankyou so much for the feedbacks, I'll pick them up in next few versions. If you try it again in a week or so, you'd find a lot of these issues resolved.

Herdr has an API btw - I am using it to build my own agent orchestrator.

I setup a Telegram bot that relays messages to the agent orchestrator that then decides "this agent needs this plugin, this working directory for context" etc to relay the work I asked. I guess it's similar to the God orchestrator.

For pure coding tasks I just want a TUI, but for non-coding work (marketing, content gen, research) an orchestrator has been great.

The pipeline + checkpointing framing resonates. The "ask me" tab needing love is a real pain point — human-in-the-loop approval gates work best when they're surfaced proactively, not buried.

Pipelines and roles are a good point.

Some more observations

- It seems like letting Michael spin up things is the prefered pathway instead of defining a bunch of agents.

- The Ask Me really needs some love. The "X/N" display doesn't make sense to me, I started on "5/5" and as I answered the questions it went to "4/5" -> .... "1/5", I'm not sure if the questions were asked in the wrong order or what. Additionally, it really needs some UI improvement for things like when there are options (and recommendations) it should show as multiple-choice (ALWAYS with a way to add an additional note, this annoys me about Claude Code today).

- Why tell the agent you spin up it's not interactive? I really think instead you should hook the AskUserQuestion and proxy that up to Michael and/or the Human directly. I _want_ to see and answer all the questions. Maybe a first pass with an LLM to see if we already have the answer?

- I still feel lost in what's happening and have to dig into agents too often to understand what has been farmed out to them. A LLM-generated summary of their task showing under their name would be nice. The Tasks board is nice but I think it and the "Ask Me" should be global tabs instead of nested under "Michael"

- Triggers seem cool, I don't have a use for them yet but I think this Claw + Agent Harness path is the correct one.

- I keep needing to unstick agents, like Michael, for things asked in the terminal that didn't get put into "Ask Me" or alert me in any way

Semi-related, I'd love some kind of SDK/library to abstract the "control XZY Agent" while being able to expose the raw agent to make writing agent harnesses like this easier. I've been playing with a little agent loop but I often wish I could have it spin up/manage a Claude Code instance and control it (I know it's possible but it's finicky).

Reply on news.ycombinator.com