It can, but I remain deeply unconvinced that the sub-agent architecture works as well as advertised.
The trick with any layering like this is that your end-to-end reliability is subagent_reliability * routing_agent_reliabilitty. Neither are 100% (or anywhere close to it, let's be honest), so the multiplying probabilities are still going to trash your performance.
If you get routed to the correct subagent, then subsequent performance is likely to be solid - but that's because you've taken the `routing_agent_reliability` term out of the equation.
Routing agent reliability hinges pretty heavily on the subagents themselves and how semantically or linguistically similar they are. If you have subagents that are in wildly disparate domains it may work well, but if your subagents start overlapping (or just look like they overlap) then routing accuracy is likely going straight into the dumpster. And a mis-route is catastrophic in that setup.
For very specific agents (well-established workflows that cross multiple, well-defined, non-overlapping domains) the architecture may be suitable, but in terms of the holy grail of the omni-agent (i.e., a desktop app agent suitable for general use) I suspect we'll continue running into a brick wall.
Reminds me of the account of the macrobians by Herodotus:
> and after this [the Persian spies] saw last of all their receptacles of dead bodies, which are said to be made of crystal in the following manner:—when they have dried the corpse, whether it be after the Egyptian fashion or in some other way, they cover it over completely with plaster 21 and then adorn it with painting, making the figure as far as possible like the living man. After this they put about it a block of crystal hollowed out; for this they dig up in great quantity and it is very easy to work: and the dead body being in the middle of the block is visible through it, but produces no unpleasant smell nor any other effect which is unseemly, and it has all its parts visible like the dead body itself.
OT wrt. TFA, but per your comment--for anyone who hasn't read Herodotus's Histories, do yourself a favor and pick it up. Great storytelling and thought provoking throughout.
I think my perfect phone would be if Apple chopped an inch or an inch and a half off the bottom of the iPhone 12 mini and ran a slightly improved watchOS on it.
Fun! Over a decade ago, my boss and I added a secret knock using Shave and a Haircut just like this to our wayfinder product, a large touchscreen app used on trade show floors.
It’s how we got into our admin to maintenance them at the shows and was useful if something was wrong with the onscreen keyboard (or we didn’t have our own keyboard to attach).
I used it for a silly credits screen in football trivia game - summer 2002 I could knock on the screen of a many pub trivia machine to make it show my name :)
I had an issue, but I figured it out. The rhythm is not the same as the one on the wikipedia page - which I was already familiar with. After looking at my recordings I realized I was playing it straight, but the default knock recording has a swing/triplet feel.
straight:
DAA DA-DA DA DA, DA DA
swung:
DAA DA D-DA DA, DA DA
I hope my highly rigorous notation is helpful here
edit - in hindsight this has nothing to do with your problem
I couldn't get it to pass on my phone, but "scram"/"get out of here" is equivalent to a password rejected/401 response. A single knock does not match the expected pattern, so you would expect it to be rejected.
Same. I've imagined 'spines.com' for books, where the work of linking book spines to ISBN's, etc, has been crowdsourced and you can point your phone camera at a shelf and look up reviews, etc.
When I visit spines.com it does a redirect to a self-publishing tool called booxAI.com . I don't think that's what you were referring to, or did the spines tool sell their domain name?
Apple iOS does this, kindasorta, but it's not real time. It does text recognition on text on images in your photostream, and you can search for text in your photos using the search bar.
Apple lidar is absolutely suited to this, and I’ve always thought there is an opportunity to integrate it into the dollhouse addon in Homeassistant.
Or, create a digital twin of your garden, and simulate light shadows throughout the day after adding or removing a tree. Add pruning schedules to a fruit tree.
It’s trivial to do a scan, but it hasn’t really taken off in a practical sense.
I would try to identify an ecosystem where he could start building on his own: Custom Shopify functionality, Wordpress plugins etc. he can try to find small customers with simple needs and scale up as he learns. The ecosystem will provide him some guardrails and shrink the surface area to something manageable. You could coach him as he goes if you have time.
I think this is a really cool idea and something I have personally wanted (or at least thought I wanted) for a long time. In fact I wrote about a similar idea in 2013.[1]
To me, this is just one part of something bigger though. I want to be able to store code in blocks and compose them into bigger blocks by chaining them and pipelining them. I'd also like versioning and some access rules.
I know this is a webdev project anyway, but I imagine embedding code would mean, in practice, embedding Javascript. Then the whole thing becomes 100% tied to the web and to modern browsers, which is a shame. Such blocks could be used from native applications, bots, or TUI programs, but that ship has probably sailed already.
We're focusing on the browser context at first, but expect the principles of the interface between a block and application (if not the implementation details) should be applicable to other environments - very briefly mentioned here https://blockprotocol.org/docs/faq#what-about-non-web-contex...
I'm interested in hiring people who want to not just invent but put into production a better way of doing that. Mostly I want people who think of the problem like that.
They say they’re are for preserving and managing context and I’ve been wondering if they help with the “too many tools” problem.
https://docs.anthropic.com/en/docs/claude-code/sub-agents