Static Harnesses Fail in Complex Real-World AI—Adaptive Engineering Needed

Rajiv Chandegra argues that AI engineering relies on fixed harnesses—predefined roles, tools, and sequencing—that work for deterministic tasks. As AI enters real-world multi-agent scenarios, this factory model becomes brittle, suppressing the variance needed for novelty.

He introduces ‘adaptive engineering,’ where engineers design constraints instead of specifying roles. Agents self-organize; the harness emerges mid-engineering, becoming the output. This horizontal coordination among agents is, he claims, more leveraged than vertical model improvements.

The approach is speculative, lacking production evidence. Chandegra acknowledges failure modes: suboptimal attractors, monoculture from shared training data, loss of legibility, and unpredictability, making traditional auditing impossible.

Engineering leaders must distinguish complicated from complex problems. Complex systems need probing, not deterministic planning. Applying factory harnesses to complex AI is a costly category error; the talk envisions self-organizing architectures, though tools are lacking.