Because an agent's behavior space can't be scoped the way a fixed UI can, Linear's quality bar for their production agent isn't "eliminate all unintended behavior" — the conventional QA standard for software. Instead they guarantee reliability only for a small set of "hero" paths (e.g., create a ticket, manage tickets, create a PR) and explicitly tolerate variance everywhere else. This is a deliberate departure from traditional QA thinking, and it's what Eval Investment Triage is built around: put eval investment behind the hero paths, not behind the full space of possible agent behavior. See also Agent Output Must Be Filed Into the System of Record for one property Linear treats as part of the guaranteed bar.
Linear names its own hero set explicitly: create a ticket, manage tickets, and create a PR. These are the heavily demoed, core paths that the team makes 'rock solid,' while accepting that behavior outside this set will vary — variance treated as inherent to current agent technology rather than a bug to eliminate everywhere. This selective hardening is what makes two-tier-model-routing viable: known hero paths get narrow, optimized prompts, and the general model absorbs everything else.
Underneath the "smart autonomous agent" framing, the actual mechanism for hitting a reliability bar is mundane: a small router sends roughly 80% of early usage to one narrow, heavily-optimized subprompt (issue creation), rather than to open-ended general reasoning. This is consistent with the broader hero-use-case philosophy — reliability comes from concentrating engineering effort on a small number of guaranteed paths, deterministic routing included, even though the public narrative emphasizes autonomous general intelligence.