Trust is the bottleneck, not model accuracy. That's the clearest takeaway from Brendan O'Leary's analysis of 40 trillion tokens of usage data, and it's a point that every engineering leader building agentic workflows should internalize. O'Leary's four-rung adoption ladder reframes the challenge: teams don't fail because their models aren't smart enough; they fail because they haven't built the context construction, multi-model routing, and feedback loops that make intelligence trustworthy in practice. This is precisely the kind of thinking that separates demos from production deployments.
The failure points O'Leary identifies resonate deeply with what we're seeing across the ecosystem. Context construction, for instance, is where most agentic workflows stumble, not because the model can't reason, but because it doesn't have the right data in the right structure. This connects directly to the challenges explored in CedarDB unifies operational and analytical data as AI agents demand more, where traditional database architectures struggle to serve both operational speed and analytical depth. If your agent can't pull the right context in real time, trust erodes before the first action completes. Similarly, O'Leary's emphasis on feedback loops mirrors the architectural evolution discussed in Exploring the Next Wave of AI Architectures Beyond Transformers, where new designs prioritize iterative refinement over static inference. Trust isn't a one-time verification; it's a continuous process of validation and correction.
What O'Leary's data makes plain is that scaling agentic workflows requires a shift in engineering priorities. Instead of chasing the next benchmark leaderboard, teams should invest in observability that surfaces where trust breaks down. Is the agent retrieving stale context? Is it routing to the wrong model for a given subtask? Are users unable to correct its outputs without starting from scratch? These are the questions that determine whether an agent moves from prototype to daily tool. The 40 trillion tokens aren't just a scale metric, they're a signal that the patterns of failure are now measurable and, crucially, addressable.
The practical consequence for engineering leaders is straightforward: build your trust infrastructure before you scale your model infrastructure. Start with a clear audit of where your current agents fail, map those failures to O'Leary's four rungs, and prioritize fixes that close the feedback loop. The open question that remains is whether the industry will invest in the boring plumbing of trust, context pipelines, routing heuristics, human-in-the-loop correction, or continue chasing model improvements that don't translate to reliable outcomes. The data from 40 trillion tokens suggests the answer should be clear.