Technical Context
I wouldn't immediately label this as a model feature. From what I see in discussions, the problem seems more like an agent configuration effect: skills like superpowers, AGENTS.md, system instructions like "don't make mistakes", plus an emphasis on tests and verification. In such a setup, Sol easily slips into an endless self-check loop.
For those doing AI integration or production AI automation, this is a familiar trap. The model doesn't just solve the task; it starts servicing its own control process: writing tests, rechecking tests, rebuilding plans, reverting steps. And at this point, I usually don’t blame the model first.
Importantly, I haven't seen an officially confirmed bug description of this exact behavior in Sol. But there are user signals that after removing everything from the configuration that nudges the agent toward planning, verification, and "perfect" execution, it begins to work noticeably smoother and stops wasting time on synthetic checks.
And yes, there's an unpleasant edge case here: the agent can delete or roll back its own work if it considers the result "not good enough." That's no longer just thoroughness; it's an architectural risk, especially if you have overly broad filesystem permissions or auto-apply changes without a human stop.
But the picture isn't black and white. There are also opposite reports: Sol handles long tasks, step-by-step migrating a project for dozens of hours, thoroughly checks, and ultimately brings the work to 80-90% of the desired state. So the same tendency to overcheck in one environment kills throughput, while in another it yields almost surgical precision.
Business and Automation Impact
For businesses, the takeaway is very grounded: don’t evaluate an agent by the "raw model." Look at the combination of prompt + tools + permissions + stop conditions. That’s where you decide whether you get useful AI implementation or an expensive self-check cycle.
The winning teams are those that limit the agent’s freedom where it’s not needed: iteration caps, separate modes for checking and for changes, mandatory checkpoint before file deletion. Losing teams are those that give the agent full access and hope that caution alone equals quality.
At Nahornyi AI Lab, we typically fix such bottlenecks not with "prompt magic" but with proper control circuit architecture. If your agent also hangs on checks, rolls back code, or performs unstable migrations, we can calmly analyze your scenario and build AI solution development so that automation with Sol actually saves time, not creates a new class of incidents.