THE ANATOMY OF A PRODUCTION HARNESS (COMPONENT MAP)

EXAMPLES Example 1: The instruction manager at work. A company has a global rule: "Always be polite and cite your sources." A particular project adds a rule: "Never modify files in the secrets folder." A user's request adds a task: "Summarize last quarter's sales notes." The instruction manager combines the three into one assembled system prompt, with the global rule first, the project rule second, and the task last. If the project rule had contradicted the global rule, the manager would apply a priority order decided in advance, such as "project rules override global style rules but never override safety rules." The model never sees three documents with unclear authority, only one coherent set of instructions. Example 2: The model adapter in action. A team builds its harness to use an internal message format that has a role and some content. For a cloud provider, the adapter wraps these messages in the form that provider requires and attaches the secret key. For a model running on the team's own computer, the adapter wraps the same messages in a different form and sends them to a local address. When the cloud provider raises its prices, the team switches the adapter setting, runs its test tasks again, and compares the results. None of the rest of the harness changes. Example 3: The permission resolver in action. An agent is working on a software project. It asks to read a file, and the resolver allows it immediately because reading is permitted everywhere in the project. It asks to write a new file in the working folder, and the resolver allows that too. It asks to push its changes to the main branch, and the resolver replies, "Approval required," pauses the run, and sends a notice to a human reviewer. The reviewer inspects the changes and approves, and only then does the push occur. The model's wording was never part of the decision, only the written rules. Example 4: The budget tracker halting a runaway. A research agent is given a budget of two dollars in token costs and twenty minutes of clock time, whichever comes first. It begins reading sources and taking notes. At fifteen minutes, with one dollar and eighty cents spent, the tracker issues a warning, and the context builder adds a short note telling the model that it is nearly out of budget and should wrap up. At two dollars the tracker declares the budget exhausted, the loop stops, and the reporter delivers a partial report that lists what was found, what remains unfinished, and a suggestion for what to do next. Without the tracker, the same agent might have kept going for hours and cost far more.