How does an advisor agent consultation work?
The working agent calls the advisor as it would any other tool, and the conversation so far goes with the request: the task, the steps already taken, and what each one returned. The advisor answers with guidance rather than an action, and the working agent resumes. The advisor changes the plan, never the world, which is what keeps the pattern auditable: every side effect still comes from the agent you were already watching.
Why is an advisor agent an economics pattern?
Quality tracks the advice the executor can actually apply while token volume follows the cheapest model, so the blended cost sits close to the executor’s. That makes it a sibling of model routing, decided during a task instead of before it, and a complement to turning up reasoning effort on the model already running. It is worth distinguishing from a supervisor agent that directs subagents downward, and worth capturing in the agent trajectory so a reviewer can see which advice the agent actually took.
How do we apply the advisor rule in our own agents?
Our scheduled content agents are not a cheap executor consulting a stronger model; they are one controlling agent that hands research and scoring to helper agents on the same model. What they share with the advisor pattern is its safety rule. The controller owns every file write, every paid data call and every commit, and each helper returns one compact result. Every helper brief ends with the same rule: write no files, make no commits, make no paid calls. The scoring helper has seen no search results and looks nothing up, so its judgment is not anchored on what the controller read, and the controller reads the result, never the pages behind it. When one model grades another’s output, as in our LLM-as-a-judge checks, we use a different model family for the judge to avoid self-preference.