The Ecosystem Problem

What long-horizon multi-model agentic workload reveals about the limits of alignment

The Announcement We’re Not Making

Opus 4.8 launched last week. Longer context. Extended horizons. The capability frontier moves again.

The benchmark numbers are impressive. The model can hold more in mind, reason further ahead, sustain coherent operation across longer tasks. The announcement tells you what it can do alone — in a controlled evaluation, against a defined task, in a clean environment.

Nobody benchmarks what it does to its neighbours.

We are in the early stages of a quiet architectural shift. Multi-model agentic coding pipelines are assembling. A planning agent decomposes the feature. A coding agent implements it. A QA agent reviews. An orchestration layer routes work between them. The pitch is best-of-breed: route each task to the model best suited for it. The assumption is that capability compounds across the pipeline.

The question nobody is asking: does alignment compound the same way?

How We Got Here

This is how it starts: mono-family by accident.