From building with agents to building the systems around them

The clearest signal from this year’s AI Engineer World’s Fair, in Latent Space’s recap, is a shift in where the engineering happens: from building with agents to building the systems around them.

Richard MacManus frames it through two essays by Lilian Weng. Her influential 2023 piece described the anatomy of an agent — planning, memory, tool use. Her 2026 essay, “Harness Engineering for Self-Improvement,” argues that the system surrounding the model now matters as much as the model: the harness that manages workflow, context, permissions, evaluation, and persistent state. That matches what we see building our own tools. The interesting work is rarely the clever prompt; it’s the scaffolding that makes an agent reliable, observable, and safe to hand real work to.

The craft has migrated to the edges — the harness, the evals, the context management — and that’s where a durable engineering discipline is quietly accumulating. Worth a read if you’re trying to place where the practice actually stands in 2026.