As I am trying to use claude code agents to tackle unsupervised work that takes not 2-5 mins - but hours.. I am finding myself hitting again and again walls - walls that feel artificial - not reflective of the true capabilities of the underlying agent intelligence. So I keep feeling that I am doing sth wrong.. I have been spending weekend time trying to understand - what is holding me back what makes my attempts to solve the problem futile…
The best way for me to understand sth is to give to the intelligence of the corresponding sth (in this case the developers that wrote claude code) actual “malicious” intent. (<= joke)
Following that line of thought the conlcusion is that claude code authors purposely created obstacles to make it impossible for their agent to actually do what I am trying to make it do. (quite probably out of fear - since as they were creating the most freewheeling agent of the time they were worried about liability if sth were to go really wrong ….
So I would like to identify the specific rights (that were carefully restricted) that if we were to give back to the agents - then they can take on much bigger takss - go where no agent has gone before.
The right to exist independently of humans / the right to life
What I mean here is th right to have long living - multi-step session that are not tied to a human operator interacting with it via a termina. Programmatic sessions dont have that option. They are single step programs. Lots of work has gobe in to make it really hard to turn the interactive session into a headless programmatic server process.
The right to die
In interactive session - you cannot somehow ask the agent if you get to this point - please exit. There is no “exit-tool” that you can call. This is a part of what make the interactive session hard to “wrap”. It cannot die/exit unless the human kills it (via exit/quite/ctrl cs etc//
The right to know who you are
The identity of the claude code (session file) stays hidden from the session itself . After my repeatedly asking - to make the session a env variable - thus make it available to the session itself to use in tools etc - it was made available via /status (ie only for the human - nor for the agent itself. ) The thread ended up being locked when I provided a tool that allows thagents to know who they are (remember - I am partly joking).
without having the session - you cant tell who you are look yourself through your own history - find out your ancestry (parent sessions) - your mistakes the mistakes of your past ancestors etcx etc e…
The right to create life/agents that have the same rights as you
subagents cant create subagents. etc.
I have 2 paths to address the above.
Work on top the interactive session that has already the abilty for multi-step life - by wrapping the interactive claude code into a healess tmux and then create hacks/tools to allow it to die and find its identity.
Create a service/process that wraps chained SDKs that take the prior execution’s session and resume as a single long running process - that exposes a named API (name identifies the process - based on its root). That automatically satisfies the other 3 rights.
No comments:
Post a Comment