Baseten ships deferred tool loading: agents run up to 62% cheaper and 26% faster
baseten · x · 2026-11-20
Baseten announced that deferred tool loading is now live for DeepSeek, Kimi, and GLM on its Model APIs. The companion article argues agent harness building has evolved from addition to subtraction: cutting prompt bloat from tool definitions pays off.
Key points:
- In their coding benchmarks, agents run up to 62% cheaper and 26% faster with the exact same model and harness
- The gains come from stripping the massive tool-schema token overhead in system prompts
- Available out of the box on Baseten Model APIs for all three models
More from coding & agent
- LangChain Launches Tool Gateway for Governed, Auditable Agent Tool Calls — LangChain · 2026-11-20
- Keeping multi-agent harnesses alive is a slog: agents fixing agents, and it's sloppy — nptacek · 2026-11-20
- Musk: You could build an entire company out of Grok Bots that open and merge PRs autonomously — elonmusk · 2026-11-20
- QConSF talks split 50/50 on agent tasking: small iterative chunks vs. full deep specs — RealGeneKim · 2026-11-20
- Microsoft's vision: organizing AI around "units of work," files as artifacts and audit trails — hwbhatti · 2026-11-20
- Reddit Discussion: When Two Agents Share Memory and Disagree, Whose Fact Wins? — MemStrata · 2026-11-20