A fundamental limitation of third-party harnesses is that they can't see all the data. As earendil pointed out, providers are now moving to encrypted thinking blocks. As an example, this could affect compaction. If pi calls another agent to do the compacting, that agent can't read the encrypted thinking blocks. In contrast, if codex calls another openai-codex agent to do the compaction, they can feed that agent the t…
The AI boom feels different from the SaaS boom though because of all the different flavors of open that are available to people right now, from self-hosting your own stuff if you're a rich tinkerer to hooking up to OpenRouter or another service and getting your pick of models. I feel like some of the biggest providers are going to find themselves on the wrong side of the moat because, despite being massive innovators in their field they're monetizing it like they did in 2014.