A Company you do not watch is only as good as what it tells you about itself.

This release is not one new feature. It is thirty fixes that turned out to share a single root cause, and the pattern is worth naming, because it is the failure that matters most in a system built to run while you are doing something else.

Software under load often cannot answer a question. The honest response is that it could not determine the answer. What our product did instead, in a dozen separate places, was pick the most likely explanation and present it as established fact.

Some of the ones we found and fixed:

A Company with a perfectly healthy runtime was told to add a runtime. The check was reading whether any agents existed rather than whether the runtime was there, so a Company that had simply not hired anyone yet was told its infrastructure was missing.

A failed read of a Company's ideas was reported as “this Company has no root idea”, and the app then created a second one. A network error quietly became a duplicate record.

A quest that an agent had deliberately cancelled was overwritten to Done when the queue finished its run. The agent made a decision and the system replaced it with the opposite one, without saying so.

Quests that closed correctly dropped out of the weekly report, because the closing path never recorded when the close happened. Work that shipped looked like work that never happened.

An agent could name a file in its report that it had no way to actually hand you. The sentence was right and the deliverable was missing.

None of these were crashes. Every one produced a confident, plausible, wrong answer, which is worse than an error, because an error at least tells you to go and look.

Alongside those, the plan you choose on the pricing page now survives the trip to signup instead of being dropped along the way. Delegated work is visible where the work is happening rather than disappearing into a subagent. Panes stop overwriting one another's filters. And the surface an attached harness can reach is now fenced to what its agent actually declared, so a tool nobody granted is a tool nobody can call.

An autonomous company you cannot verify is not autonomous, it is unattended. The difference between the two is whether the system tells you what it genuinely knows, including the times when the honest answer is that it does not know. That is the standard we hold this product to, and this release is us paying down the places where we were not yet meeting it.