It looks like Cowork can totally corrupt its session if it hits API errors while working, and the state is totally unrecoverable.
Anthropic's products are genuinely impressive and their rate of delivery is admirable, but bugs like this—especially in the AI space—always feel comparatively disconcerting. A bit like if I noticed medical equipment in a hospital clearly malfunctioning. Aren't we supposed to care more?
I know Cowork isn't a mission-critical product, that comparison is unfair, and logical issues in it aren't likely to threaten human life or anything like that. Just, the slop-factor in an organization where I'd hope to see distinctly less of that is disconcerting. Sometimes I sit and watch the Claude Code TUI flicker and think wow, this is a solved problem. This is weird. But it does get the job done.
It must be really hard to balance quality with MF&BT at these companies right now.
I should note: I'm sure I'd deliver complete shit if I were building Cowork; not criticizing them or their abilities so much as lamenting this strange aspect of AI development in 2026.
