THE SEVEN-LAYER ETCLOVG TAXONOMY

EXAMPLES Example 1: The Execution layer as a loop. At its simplest, Execution is a while-loop, meaning a piece of code that repeats as long as a condition stays true. In words, the loop says: keep calling the model until it says it is done or until we have repeated too many times. Suppose an agent is asked to find the cheapest flight. On the first pass the model asks to search for flights, the harness runs the search, and the loop goes around again. On the second pass the model compares prices and says it is finished, so the loop ends. If the model had never said it was finished, the step limit would eventually stop the loop. That tiny cycle of calling, acting, and repeating is the heart of the Execution layer. Example 2: A tool with a defined shape. Imagine a tool called search_web. Its input schema, which is the agreed shape of what you hand to it, says: give me one piece of text called query. Its output schema, the agreed shape of what comes back, says: here is a list of results, and each result has a title, a short snippet, and an address on the internet. Because the shapes are fixed, the model knows exactly how to ask, and the harness knows exactly how to read the reply. Defining those shapes, and deciding which tools the agent may see, is Tooling work. Example 3: Two teams, two concerns. On Monday, the reliability team notices on a dashboard that the average time per agent step has doubled this week. That is an Observability finding, and the team begins investigating whether a tool has slowed down. On the same Monday, the security team reviews an access report and discovers that one agent run opened a file in a restricted folder. That is a Governance finding, and the team begins investigating whether a permission rule is too loose. The two teams look at different data, ask different questions, and may even disagree, for instance when the security team asks for logs to be shortened while the reliability team wants them lengthened. This is exactly the situation the separation is designed to handle. Example 4: One support run touching all seven layers. A customer writes, "My order arrived broken, please refund me." A support agent handles the request. Execution runs the loop that thinks and acts through the task. Tooling lets the agent look up the order in the order system. Context keeps the customer's message and the order details in front of the model. Lifecycle starts the run with the right settings, enforces a time limit, and ends the run once the case is closed. Observability records every step so that a manager can later see what the agent did. Verification checks that the proposed resolution matches the refund policy and scores whether the outcome was correct. Governance checks that this agent is actually allowed to issue a refund of that size, and if not, requires a human to approve. A single ordinary request, then, passes through every layer.