Kilo Code's gateway talk ties AI agent adoption to a trust ladder

The ladder is a dependency chain. O'Leary describes autocomplete first, then chat with agents in code, then orchestration of many agents at once ^[11]. His claim is that a failure low on the chain blocks everything above it. Slow suggestions, a wrong feeling, or an agent that edits the wrong file are his examples, and in his words the upper rungs "really don't get a chance because the trust is broken" ^[12].

He then names three places where adoption breaks: how context is constructed, how models are routed, and what feedback loops exist. He says the data shows a consistent pattern at each and that each has a concrete fix ^[13]. We think this is the part of the talk most likely to transfer. All three are properties of a team's own harness. They do not depend on who the users are.

The usage numbers are harder to carry over. O'Leary works for an open-source competitor, and he put his position this way: he sits at "an uncomfortable intersection" between what Kilo can see and what larger vendors say ^[5]. The data comes from Kilo's gateway, which sits behind a CLI, a VS Code extension and cloud agents [c2, c3]. For a figure from that gateway to match your rollout, your developers would have to reach models through similar surfaces. They would also have to resemble people who chose an open-source agent tool in the first place.

His case against benchmarks is anecdotal so far. He describes the industry story as capability: one model beats another on a benchmark, another has a bigger context window ^[7]. He says a gain of "0.02 points" on human eval has "very little to almost nothing to do with whether developers can actually adopt those models" ^[8]. As evidence he offers what he has seen: "I've watched teams get access to the best models, the best capabilities, best benchmarks, and not use them at all" ^[9]. He puts the gap down to trust ^[10].

The headline figure carries its own caveat. Of the title's 25 trillion tokens, O'Leary said: "We'll see in a minute that's not quite right" ^[2]. The excerpt's outline ends at the third point, trust, before any chart or count ^[6]. We can report the framework. We cannot report what the gateway data shows.