Tool result offloading keeps large outputs out of the model's context

Tool result offloading writes an oversized tool result to storage instead of into the model's context. Set offload_tool_results=True and any result over 16,000 characters goes to AgentFS, leaving a short envelope in the message with a preview, the size, and a result_id. The agent gets read_result and search_result to fetch the rest only when it needs it, and nothing calls a model on the write path.

Previously, a large tool result landed in context in full and stayed there, consuming your window and adding tokens to every turn that followed. Now the agent carries a compact reference by default and pulls the full result, or searches inside it, on demand.

You can tune when offloading happens and how long results are kept with ResultStore(threshold_chars=..., ttl_seconds=...).

Turn it on when you create an agent:

from agno.agent import Agent
 
agent = Agent(offload_tool_results=True)

View the Result Offloading docs. See the cookbook for reference.