Are AI coding agents hitting a wall, or are we just measuring them wrong?

I keep seeing three AI-agent conversations collide: Big companies saying agent progress is slower than expected New benchmarks trying to judge agents like senior engineers Builders saying agentic coding works, but only with strong constraints, logs, reviews, and cleanup My current read: the hype made it sound like agents would replace engineering work. In practice, they seem better at accelerating parts of the workflow, but only when a human still owns architecture, context, debugging, and review. The uncomfortable part is that agent failed often means the workflow around the agent was vague. So I’m curious: What actually works for you in production or serious side projects? Where do agents still fall apart? Are benchmarks like SWE-Bench measuring the right things? Is the real skill now coding, or directing/reviewing agents well? I’m especially interested in examples beyond toy demos: multi-file changes, debugging, tests, refactors, dependency choices, and long-running tasks. Are agents underperforming, or did we expect autonomy before we built the guardrails?