Deferred tool loading: Make your agent 62.5% cheaper without changing your model or harness
Deferred tool loading: Make your agent 62.5% cheaper without changing your model or harness
Baseten added deferred tool loading to open models (DeepSeek, Kimi, GLM) on its Model APIs, supporting Chat Completions and Messages API formats. In Claude Code with three MCPs enabled, coding benchmarks were on average 62.5% cheaper and 26.2% faster, while agentic benchmarks were 51.1% cheaper and 14.9% faster, with tool schemas dropping from about 141k tokens to about 1.3k. The technique keeps tools out of the system prompt and gives the model a tool_search tool that renders schemas inline at the bottom of the transcript, preserving the KV cache.
It offers concrete benchmark numbers and a clear mechanism for cutting agent cost and latency without changing the model or harness.
Source: Baseten Blog · baseten.co