PwC looked into Anthropic's harness primitives (a loop with memory, filesystem, and bash, i.e. the Claude Code shape) and argues that enterprises should stop building a new agent system per use case and instead deploy one identical harness everywhere.

This happens mostly because it collapses review, not because it runs better.

So they recommend standardize the loop across teams and let the prose in an instructions file carry the difference.

If every agent in a company runs the same unmodified harness and differs only in a plain-text instructions file, auditing the whole fleet becomes reading documents instead of reviewing code.

– arxiv. org/abs/2608.20622

Title: "Applying Anthropic Primitives at Large Enterprises: Harness Paradigm for Knowledge Work"