I'm also curious, if you just ask Codex, Claude etc already do subagent management on their own, and if instructed, can certainly set up a structured communication system for the group. Why these Herdr, Pi etc products on top?
Is Codex, Claude Code mismanaging when no tokens left? Or any other specific reason?
> Codex, Claude etc already do subagent management on their own
What about together? Can you reasonably do agent to agent interactions across harnesses? herdr makes that easy.
One of my favorite reasons for using herdr is what I'll call "everything is a pane". Running a new agent? Runs in a herdr pane. sshing into a system? Runs in a herdr pane. Anything that runs in a herdr pane, your agents can interact with. I will often do things like ssh into a system in one pane, and then ask pi in another pane to help me do diagnostics on that system. It never has to try to ssh in directly in the harness with my credentials, it just sends keyboard input to the pane, and reads the content of the pane. I can watch the interactive session as it goes by.
Another workflow that I have sometimes is I'll kick off some work in an agent, and instead of making a new sub agent, I'll actually make a new parent agent if I get worried that it's going down an unnecessary rabbit hole. I'll tell the new agent something like "periodically watch the agent in pane X, it is working on Y but I am concerned that it is losing track of its tasking. Periodically review what it's doing, and make sure it doesn't go off the rails. If it does seem to be going off the rails, give it guidance to bring it back on track" and that seems to have helped.
Subagent spends more tokens. It can be just separate sessions, just a new codex or claude session it doesn't have to be subagents that have to coordinate
Nvidia is investing $1 billion in Poolside and paying $6 billion to license its technology and hire most of its engineers.
They are the elephant in the room, with their CDS sky-high. Sounds like they are opinionated on the whole AI thing, trying to drive this with open-source models, countering OpenAI using Jalapeno.
Looks zero-sum for the players.
All the while Google silently planning to get milk from all layers.
They do reach out pretty quick when your usage takes a huge dive after you move to one of the single provider plans that offer hard to believe possibly unsustainable value.
"Google" doesn't understand anything, as it isn't one being.
Google has also as many average employees as any other company, and they will also come up with bad, corporate, ideas.
Honestly the worst part about the LLM age is that everyone is suddenly an "expert", and that is why we get shitty things like A2A or MCP or whatever the next "shiny" overengineered thing is.
Just like with any new technology, people will crawl out the woodwork to establish "standards" just so they can claim fame, money and attention that comes with it.
A2A is a very high level protocol that's meant to be an entry point for the actual network protocols that allow agents to communicate with others like HTTP, gRPC, and others. The promise of no-code means the concept has to exist as an abstraction.
There's a Kaggle course going on where Google discusses A2A and their thoughts on it. If you're a pure vibe coder who doesn't know how to code, A2A is for you. If you know the basics of programming and could even do some web crawling, there are many accessible options.
Google and the industry is honing in on these vibe coders who will look at 10 million tokens consumed to make a checklist application and think nothing of it. The agent to agent (A2A) protocol is for them. Personally, I think it's useful to describe what I'm already doing to people who aren't experts.
Full wave inversion uses all of the information from the wave and more intense computational tomography to image structures that pulse wave B mode cannot, though gases are still a problem. Computationally, if you squint, it's similar to the work Midjourney does with AI image generation, as it progressively generates a structure that fits the data.
Ultrasonic waves can penetrate most structures in humans, including the brain. For example, with focused ultrasound (as they mentioned with MRgFUS) you can burn specific structures in the middle of the brain without any incision.
To use this for imaging, you need lots of transducers (MRgFUS typically uses 1024 for ablation, and Midjourney is proposing 358,000 for imaging) and massive advances in computational tomography capabilities. There will still likely be pockets of low confidence where there's a lot of air, like in the lungs. But with sufficient information on what's happening around those areas, you'd still have something that's medically useful.
Ultrasound researcher here: this work is almost certainly not full wave inversion. You're right that FWI can be done in this kind of setup. Notably, a lot of people are trying to use circular arrays of ultrasound sensors to do FWI, particularly in the brain or the breast. But, it's a challenging inverse problem to solve and as far as I know, it's still extremely slow - nowhere near realtime.
The data they show is almost certainly a normal beamforming algorithm like delay and sum, possibly with some simple speed of sound correction. The most similar paper I know of is here in Nature Biomedical Engineering from a team at Caltech. https://www.nature.com/articles/s41551-026-01660-4
Is Codex, Claude Code mismanaging when no tokens left? Or any other specific reason?
reply