The Software Around AI Shapes What It Can Do

An AI product's abilities depend on more than its written answers. At its centre is an AI model that responds to requests. An AI model is a system that produces answers from information. But the model cannot check a calendar or send an invitation alone. Other software must connect it to tools that perform those actions. The article calls this surrounding software the agent harness.

The harness manages the information, actions, and limits around the model. It checks a request, uses a tool, and returns the result. The model can then ask for another action or answer the user. Two products using the same model may therefore work very differently. One meeting assistant might only improve the words in an agenda. Another could check calendars, find free time, and send invitations.

The surrounding software also passes useful information to the model. This can include earlier messages, work hours, and the user's saved choices. A saved choice does not help if the model never receives it. The software also controls which actions need the user's permission first. Reading a calendar and creating a meeting can need different permissions. Approval in one part of the system cannot remove another service's restrictions.

Teams must also decide when the AI should stop or ask for help. A meeting search may find no time that works for everyone. The product could explain the problem or ask about another week. Testing should cover the whole task, including these difficult situations. A well-written invitation does not prove that the calendars were checked. Teams should also record how much checking and fixing users must do.