The useful part of AI coding agents is not that they can type code quickly. It is that they can explore alternatives, take on bounded work, and surface implementation options a human can review.
The weak part is everything around that: context preparation, safety boundaries, diff review, verification, retry decisions, and the question of what evidence survives after a run finishes.
That is the shape of Consortio. I want agent work to look less like a long chat transcript and more like an engineering workflow, and each property earns its place by fixing a specific failure. The task is written down as durable intent because otherwise there is nothing to judge a run against. Candidates run in isolated workspaces because comparing alternatives beats trusting whichever one raced to "done". Ranking stays downstream of verification because a confident summary is not evidence. And export waits until a human can inspect what happened, because that is the moment a bad change would otherwise slip in.
The interesting product problem is not "which model is best?" It is how different agents, reviewers, judges, deterministic checks, and a human owner can work together without turning the process into noise.
That is where the interface has to earn its keep. It should show disagreement, unresolved risk, and evidence without making the user read every raw artifact by hand. The same review-first stance shapes Masterful CV, where AI proposes changes a person approves before they ever touch the document.