On the flip side if there is a bug in one of my systems I would rather get a detailed bug report from an llm than a user message “xxx does not work”.

This is so obvious that in practice it doesn’t buy much, but everyone is still propagating that silly news. This is the real malware, a mind virus.

We don’t trust llm execution- so we add user approvals. But task decomposition calls for co-recursion between code and prompts. This means that the approvals should be evocable at any depth. I think we need some kind of protocol for that (à la the Cubes OS protocols for cut and paste between vms).

Maybe a workaround could be to use bubblewrap of the scripts ther recursively call the llm (and run the agent in yolo inside the wrap).

Well, or not spawn any external commands, and actually have tools made of code written by someone who thought about what the agents at each level should be limited to doing.

Or just run agents in a container…

In the limit we want the llm to write the code (like in RLMs).

Currently, having LLM feeding on its own output repeatedly is the fastest way to get it hallucinate.

Transactional recursive agents ?

Nothing is committed until the final top-level transaction is accepted.

Too late for fixing it - but of course I meant https://www.qubes-os.org/

zerostack contains --sandbox flags that forces bwrap usage on all shell tool usage

The idea is that the spec is somehow compressed in comparison to the code.

You need supporting environment on both sides of the sandbox.

Frontmatter should now be in the markdown standard.

I have been struggling with that. Thanks! Let me reword it - natural language lacks a strict semantics - so also programs for the llm machine ( I.e. prompts) cannot have it. LLMs always have to project from all possible semantics into one (are there any experiments with superpositions?)

What is the root cause of this? It is said that AMD is hardware company and neglects software - but recently they issued lots of declarations of becoming software firs now.

The point is for the agent to have an index of available skills so that it can decide autonomously if it needs to load one.

https://zzbbyy.substack.com/p/what-are-skills

In my experience it doesn’t work too well with codex, but I expect llm providers to train them on that use case and improve the situation soon.

At some point evolving software instead of designing it will work. Now the evolutionary pressure leads towards churning more tokens.