What is interesting about the rise of harnesses and manager layers is the irony of what we are building.

Take Google Cloud Platform’s Scion: an orchestration layer that runs agents in isolated containers, assigns them specialized roles, and gives each one its own configuration, identity, and behavioral contract.

We are taking state-of-the-art models and wrapping them in familiar structures of hierarchy, delegation, and control, partly because raw interaction is still too unreliable and partly because products still need auditability, control, and fault isolation. Replace your favorite harness here, whether Gastown or Agent-flywheel or Forgecode.

It is hard not to notice the irony.

As models become more capable, the control structures we build around them start to look more and more like the human institutions and suboptimizations they were supposed to simplify.

That, too, feels familiar.

Organizations often end up shaped around the working style, preferences, and blind spots of the people at the top, then wonder why they never quite become the high-performing systems they aspire to be.

Harnesses can feel a bit like that: structure layered on top of intelligence, partly to compensate for capability gaps, but often also to compensate for unreliability in how that intelligence behaves.

At this rate, we may end up with a bicameral system of governance, complete with elections and voting, just to write software.

Do I have a solution to this? No. For now, it is simply an observation.

What is especially striking is how often this field seems to rediscover coordination problems in new clothing.

TCP, CAP, distributed systems, Bitcoin, and now agentic harnesses all circle the same underlying problem: coordination under partial trust.

The names change.

The curse of Babel apparently does not.