Void Freud

Had a look at Cordis, the theory behind DeepSeek Harness's plugin architecture. Quick thoughts: It's basically a theory that, as long as you follow a set of contracts for the components of a harness, there are theoretical guarantees that your harness can safely remove a component. It can potentially be powerful, especially when you have a gigantic harness that is self-evolved. The agent might do some real-time experiments, sort of like A/B testing: adding a new component, and if that component isn't useful, reverting it. In this framework, as long as you know the contracts are followed, you don't have to reason through the complex global dependencies to figure out how or if you can safely revert it. The main limitation though, is that those constraints aren't very easy to follow precisely in practice, especially with an LLM involved. For example, constraints like module independence and acyclicity are easier to reason about in a purely programmatic setting. However, once the harness has an LLM in the loop, the LLM context window naturally introduces really complex dependencies. Essentially, everything can end up depending on everything else. The framework ofc might still be powerful for designing harnesses, something like React for harnesses. To me, it depends on to what extent the empirical robustness benefits still hold despite the contracts not being followed exactly, versus the additional work introduced by trying to follow those contracts. Either way, I think it's super exciting that there are more and more attempts to treat harness design as a serious science instead of a collection of ad hoc tricks. Very bullish on the direction of mathematically grounded harness design in general.