Claude’s new Advisor Strategy is pretty interesting

A lot of people building AI agents run into the same problem sooner or later.

If you run the entire agent on a powerful model, it works well but the costs grow quickly. If you run everything on a cheaper model, the system stays fast and affordable but it sometimes makes weak decisions, especially when planning complex tasks or choosing tools.

Anthropic recently introduced something called Advisor Strategy that tries to solve this in a simple way.

Instead of using one model for everything, the agent runs on a smaller executor model like Sonnet or Haiku. That model handles the normal workflow such as calling tools, executing steps, and moving the task forward. When the agent reaches something more complex, it can consult a stronger model like Opus for guidance. The advisor reads the full context, suggests what to do next, and the executor continues the workflow.

So most of the work stays cheap and fast, but the agent can still get strong reasoning when it actually needs it. It feels a lot like how a junior engineer works most of the time but occasionally asks a senior engineer for advice.

I found this architecture interesting because it pushes agent systems toward multi-model setups instead of relying on a single model for everything, which seems like a direction many frameworks will probably move toward.

I made a short video breaking down how the advisor strategy works and how developers can implement it in their own agents