Okay, new project. Making a agent harness. Command-line only, very limited scope, intended for some coding but mostly do writing and other tasks.
While concerned code could go rogue, it seems far less likely if it's read-only operations. Plus with memory management built-in 'compacting' is less an effect of suddenly losing mental competency and more a method of navigating summarized posts and can retrieve them if they are needed.
With compiling in mind, i could then have it attempt to compile say C/C++ code, see the errors and attempt to fix it. While executing external code (outside of select tools) would be disallowed. I am iffy about any scripts or programs running, but having seen and authorized 10 calls in a row i can quickly see it getting very monotonous and boring, and verifying the command in it's entirety is out, especially if i want it to do something overnight.
I've tried giving a command of 'write me a story of X' and then i get back the next morning and only the outline or chapter 1 was written, making it pointless to get something whole to glance over and see if it was worth it. Plus chapters tend to be.... super small. Like 200 words per chapter, hardly a page or two.
The idea of having the harness given multiple prompts and then having it chat back and forth under different ID's seems interesting too... Possibly for story or writing purposes rather than doing it in a single output. (700-1000 token outputs is higher quality per post than say 10,000 token output trying to do the whole thing).
Well it's worth trying out. Two harnesses, one to chat under 2 different identities, and 1 for agentic semi-work...
While concerned code could go rogue, it seems far less likely if it's read-only operations. Plus with memory management built-in 'compacting' is less an effect of suddenly losing mental competency and more a method of navigating summarized posts and can retrieve them if they are needed.
With compiling in mind, i could then have it attempt to compile say C/C++ code, see the errors and attempt to fix it. While executing external code (outside of select tools) would be disallowed. I am iffy about any scripts or programs running, but having seen and authorized 10 calls in a row i can quickly see it getting very monotonous and boring, and verifying the command in it's entirety is out, especially if i want it to do something overnight.
I've tried giving a command of 'write me a story of X' and then i get back the next morning and only the outline or chapter 1 was written, making it pointless to get something whole to glance over and see if it was worth it. Plus chapters tend to be.... super small. Like 200 words per chapter, hardly a page or two.
The idea of having the harness given multiple prompts and then having it chat back and forth under different ID's seems interesting too... Possibly for story or writing purposes rather than doing it in a single output. (700-1000 token outputs is higher quality per post than say 10,000 token output trying to do the whole thing).
Well it's worth trying out. Two harnesses, one to chat under 2 different identities, and 1 for agentic semi-work...
