The agent loop is the recurring cycle: observe context, reason about the next step, act through a tool, inspect the result, update the plan and stop or continue. The exact internal implementation differs by product.
Evidence-based completion means declaring success based on inspected diffs, passing checks, rendered output or other concrete evidence, not the agent’s own confidence. The loop is only useful when observations are reliable and stopping conditions are clear.
