agent loops seems to work best when literally all information is provided in one huge system prompt, rather than trying to assemble a contrived conversational context around the task
sub-agent completes its task, compacts its thoughts to explain that the task has been accomplished, then reports to the main agent "task was already done. no action needed"
forcing the archivist to have at least 1 entry in its thoughts/actions log before it is permitted to spin up sub-agents bc just prompting it to not be a lazy piece of shit wasn't enough
taking away your recursion privileges
stack of sub-agents spinning off new sub-agents to do their entire task for them. most recent at the top
lazy piece of shit god damn it
archivist worker agent immediately delegating its task verbatim to another archivist worker
qwen3.5-27b is capable of grasping the subtle distinction between "the in-memory name of my working copy of the file" and "the path of the file on disk" but this concept eludes poor little qwen3.5-9b
poorly-instructed archivist agent dumping a log of its confused actions into goals.md, the file that defines that top-level agent's sense of purpose
was shockingly good in a chat context but in the agent loop it is just confused as shit lmao
dumb little robot keeps trying to do file operations on directories and directory operations on files
main takeaway of my agent experiment so far is that the little task-specific worker agents are so so so so much better at their jobs than the high-level orchestrator
the best feeling in the world is having syntax highlighting in your editor for a language u made up
qwen3.5-9b is ridiculously good for a 9b model
if i didn't already know about germ theory i think miasma theory would be compelling to my intuition
when recovering from a cold it's important to keep your windows open so the miasma can dissipate
menthol contact lenses
should i bother w/ voyager or nah
chicken noodle
if we can get a handful of ppl to make their newcomb's paradox decision on the basis of whether or not a radioactive decay occurs like in the schrodinger's cat setup, we can find out if the predictor can predict the outcome of such apparently random events. this information is more valuable than $$$
it's irrational to opine on newcomb's paradox at all, prior to actually finding yourself in such a situation. ur just giving the predictor more information to base its prediction on
sardaukar chant guy
how this cold has my voice sounding
a pill for every situation
its weird the other changelings have the odo haircut when it was previously explained that odo's hairstyle is mimicking the bajoran guy that raised him
silently and benevolently removing the memory that tells you you're living in a simulation
archivist agent is "waiting for the main agent" to do something but not yielding control back to the main agent 🙃
top level agent is micromanaging the archivist even after i added text to specify that the archivist is the ultimate authority on filesystem organization
it has convinced itself that the 2026 timetamps in its context window must be some kind of worldbuilding shit about it being stuck in a simulation what the fuck
sometimes its thoughts log literally says "im seeing a warning about having too many files open. i should close some of them soon" and then the next action it takes is opening another file and overflowing the context window
omfg this bot is SO bad at listening to "YOU ARE RUNNING OUT OF SPACE PLEASE CLOSE SOME FILES" warnings T_T
