Agents / Standard term
Agent harness
All the software wrapped around an AI model that turns raw intelligence into a reliable, controllable product: permissions, retries, logging, memory, state tracking, and error handling.
The AI model generates plans and text, but the harness is everything that makes those outputs safe and dependable in production. When an agent tries to call a tool, the harness checks permissions, validates the arguments, runs the call, handles failures with retries, logs what happened, and tracks the conversation state. Think of the model as a skilled worker and the harness as the safety equipment, scheduling system, and quality-control process around them.
Builder example
When an AI product fails in the real world, the problem is almost always in the harness: a missing retry that turns a temporary API error into a crashed workflow, a permission gap that lets the agent access data it should not see, or no logging to figure out what went wrong. Investing in your harness pays off more than switching to a smarter model, because reliability is what makes users trust the product.
Common confusion: A model demo and a production product are very different things. The model might ace a task in a demo, yet the product fails repeatedly because the harness lacks retries, permission checks, or graceful error handling.