Post

Thought experiment: it's 2028. Anthropic has crashed. Claude is no longer available. Can you or those on your team maintain the code it wrote? Or will you simply look for a replacement which, of course, will be able to maintain that code?

Do you think people who take a year off work (for any reason) lose all of their skills and become incompetent?

Replacement models already exist and are open weights so won’t go away. And like the dark fiber left everywhere after the dotcom crash, datacenters full of GPUs will be available at reasonable prices with all sunk costs cleared by bankruptcy

So, what you are saying is that code written by AI can and only will be maintained by AI? Does this also mean that the evolution/development of new code languages will stop?

Also, once TSMC and Nvidia margins have come back down you’ll be getting the DGX Spark successor as a NAS.

not to mention the models (whose training costs have been paid for) will also be available at bankruptcy auction

Not exactly agent to agent but I built an app that exposes an MCP and a read only user interface. An agent who can speak to the MCP can write, read and update notes maintained by the app. The motivation was so that my Claude and my husband’s Claude could collaborate on planning a trip. 👉🏻

It worked really well. To test the multi user stuff I would start two claudes. Each read a test script from the app and used a note in the app as a semaphore to coordinate their activity.

Oh … I actually opened the link! It’s exactly this use case > The example went, "I'll tell my agent that I want to go out this weekend, and then my agent will talk to all my friends' agents, and they'll help us make plans."

I think companies having some sort of communication bus that is agent only is already happening just in annoying stupid ways lol

what use case pushed you toward it? for me it was tasks with different context needs, one agent scoping the work before another executes made both cleaner

Pretty much that logic. Controlling the context window is really important

Is this just a (cleaner) alternative to having the foreground agent dispatch a subagent to trawl the knowledge base, or are there other benefits as well? I suppose it might be cleaner to have a standardized “subagent” in front of the knowledge base for all users and harnesses.

I have a Claude code skill /Conductor which allows one supervisory session to communicate with (via a file bus) other sessions. I have to set up the sessions but from there it runs autonomously. Using Fable as the supervisor and mostly Opus for the workers gets good results.

agent to agent communication may or may not be big, but I do think it's pretty handy coherence uses a decision log so that concurrent subagents can relay findings and decisions to one another but even better, these act as letters to FUTURE agents, as well github.com/daniloc/cohe...

The letter to future agents is what sticks. Concurrent handoffs help, but the same log is what a fresh agent reads days later when nobody remembers why path B died. Without it every session rebuilds intent from the diff alone, and that is where review time blows up.

Any resources/links to the Bluesky company knowledge base agent? Sounds cool and would love to learn more about it!

@thebadcode.com anything on this you can share?

working on something fit for public consumption, but heres some examples of it in action (which looks quaint for only being a month old): bsky.app/profile/atti...