Picking just one model for an agent means picking a compromise.
Go big and you pay frontier prices on every trivial step. Go small and you're exposed the moment a genuinely hard sub-problem shows up.
Our new AdvisorTools gives you a compromise-free path forward. It lets you run a fast, cheap model as the primary and it consults heavier models only when it decides it needs to.
To set up, you hand it a list of advisors. You can even add a note on what each one is good for. The agent then gets two tools. It can ask one advisor by name, or ask them all and compare what comes back.
from agno.agent import Agent
from agno.models.anthropic import Claude
from agno.tools.advisor import AdvisorTools
agent = Agent(
model=Claude(id="claude-haiku-4-5"),
tools=[
AdvisorTools(
advisors=["google:gemini-3-pro", "openai:gpt-5.5"],
descriptions={"gemini-3-pro": "Strong at long-context analysis"},
)
],
)The descriptions do real work. They tell the primary model which advisor to reach for, so a long-document question routes to the long-context specialist and everything else goes where you pointed it. The primary keeps control of the conversation and spends the expensive tokens only when a problem earns them.
Take a minute to think about what this changes about model choice, and how your work will change when you can stop betting the whole agent on one model.
See cookbook for reference.