Agent Harness

The runtime layer around a language model that lets it operate as an agent: tool dispatch, memory and state, sandboxed execution, context management, and guardrails. The field's compact form is agent = model + harness. The term is polysemous by its own literature: it can mean a whole shipped product, an evaluation scaffold, or the user-assembled outer harness of instruction files and connectors, and only the first of these is a product with a grant. The word is borrowed twice over, from tack and from the software test harness (the stubs, drivers, and environment that run code under observation), inherited into AI through reinforcement learning. Coinage is contested (Mitchell Hashimoto and Vivek Trivedy are both credited).