Agno introduced an offload_tool_results=True option for its agents. Any tool result over 16k characters is written to AgentFS, and the message keeps only a preview, the size, and a result_id. The agent can call read_result to pull the full output or search_result to find part of it, with no model call on the write path.
The mechanism stops large tool outputs from being re-sent as tokens on every later turn. The post does not report measured token savings or how often agents need to fetch offloaded results back.