Editing Openai/694057b6-101c-8007-9a65-c40578c7252d (section)

=== User: Timeouts won't need per request overrides. ===
Timeouts won't need per request overrides. Models will be setup based on the kind of jobs they are expected to handle.

The tools system is focused on lightweight tools for lower level and smaller models, larger and smarter models will be given access to heavier and more consequential tooling using the same systems. Though shouldn't change the way it is implemented beyond what helper class is plugged in, but assumptions need to not be baked too deeply into the call mechanism.

I think the BIOS should also contain basic universally applicable system prompt data, such as telling the model that it is operating as a hivemind of cooperating models, with the passed in system prompt being more flexible and changing according to the job type.

The non-goals in section 12 are all correct.

Next is coding style. Simplicity, Clarity, robustness, and a solid test suite are to be prioritized over clever tricks. As with the nuke and repave plan: the model time lost in a couple minutes of dumping a malfunctioning server and reloading is less of a problem than a system which accretes cruft.

A small caveat is that this will be working in a somewhat CPU and RAM constrained environment. It should not ''waste'' compute in it's functioning either.