The bottleneck shifted

We've invested heavily in our coding agent harness and it's gotten scary reliable. Our dev runtime is fully legible — we can prompt agents with constraints like “ensure service startup under 800ms” and “no span in these critical user journeys exceeds two seconds” and they handle it across the full stack with minimal token usage. CI is a mix of deterministic and agent-based checks, and when it fails, agents auto-fix. Code review has become negligible — if it passes CI, we're good.

Code write and review feels genuinely solved for us. That time shrinks every day as LLMs improve and the harness matures.

What's not solved: when multiple engineers are steering concurrent agents on a large feature, how do you make sure everyone's headed in a cohesive architectural direction? Our primary velocity risk these days isn't code quality — it's high-level intent mismatches across parallel workstreams. Building the wrong thing really fast, in five directions at once.

Right now the industry's answer is sharing plan-mode markdown files in Slack and hoping someone reads them carefully. The irony of having a world-class agent setup and coordinating intent via copy-pasted text files is not lost on us.

We've started calling this “intent review” — reviewing the plan and direction before agents run, instead of reviewing the code after. The bottleneck has shifted from execution to alignment, and it's an innately human problem.