Technical Context
I wouldn’t call this a confirmed platform bug at the documentation level just yet, but the case is too ugly to ignore. A user gave Sol a long report‑building task, waited about two hours, received a request to switch to a simpler model, declined, and then all progress vanished.
This is where I immediately think not about drama but about AI implementation. If the agent runs a long job inside a single session and never writes intermediate state to the outside, any downgrade or reset attempt turns two hours of computation into zero. The token limit, judging by the description, was already spent.
I have no official confirmation of exactly this Sol behaviour. Public material about Sol currently focuses on capabilities, partner access and agentic scenarios, not on what happens when you refuse a model change. But the pattern is familiar: a long agent run, internal subtasks, context repacking, then a state reset.
Technically it could look like this: the agent hits a context, cost or tool constraint, suggests a downgrade, and when refused it can’t properly preserve its working state. As a result the session memory rolls back, draft artefacts aren’t saved, and billing or the task limit already marks the attempt as completed. Yes, this is exactly the case where “almost finished” means nothing.
I would treat this not as a one‑off oddity but as a warning for everyone building automation with AI on long asynchronous tasks. If the result matters, it should be saved step by step: summaries, checkpoints, external storage, stage‑wise artefacts – not just the final chat reply.
Business and Automation Impact
The losers are those who give the agent long reports, research or multi‑step pipelines with no checkpoints. The winners are those whose AI integration is built as an engineering system, not as a bet on a single successful run.
My practical takeaway is simple. First: don’t keep critical progress only inside a Sol session. Second: before risky steps, force a brief summary of what has been done and save it outside the chat. Third: break the task into stages with separate result capture, even if the UI promises “autonomy.”
At Nahornyi AI Lab, we cover exactly those gaps for clients: we externalise state, design checkpoint logic and build AI automation so that a model failure doesn’t burn hours of the team’s work. If your own long processes are already cracking at the seams, I can re‑architect them with you so that the agent helps the business instead of running an expensive lottery.