The next form of AI vendor lock-in may come from hidden agent state rather than model weights or prompt formats. The post argues that long-running agents increasingly depend on provider-managed memory, encrypted reasoning items, and custom execution environments.
The concern is practical. OpenAI’s Responses API can preserve reasoning across turns through server-side state or opaque encrypted items that a developer can replay but not inspect or translate. Mozilla notes that Anthropic and Google use similar hidden reasoning artifacts, which means continuity built inside one provider cannot be cleanly transferred to another.
That matters for teams trying to build portable AI systems. A plain-text handoff can summarize conclusions, but it does not carry the same internal state that shaped the previous steps. Benchmarks can also become harder to compare when scores reflect a model plus a proprietary harness rather than the model alone.
Mozilla is not saying these tools lack value. It says the performance gains are real, but developers should recognize the architectural trade-off before tying critical agent workflows to state they cannot examine or move.