Read a piece on cutting costs in a Harness workflow. Noting the main points.
https://zhuanlan.zhihu.com/p/2074164605908153019
[Token costs down 50%: a cost optimization practice for Harness workflows — Tencent Technical Engineering]
Their approach comes down to three principles: let the AI see only the context it needs right now, cut the irrelevant context, cut the repeated context. Following those, they reworked ten areas — architecture decomposition, stable prefixes, progressive disclosure, code graphs, CLI instead of MCP, on-demand indexing of long-term memory, parallel tool calls, and so on. Added up, the estimated saving across the whole flow is 50%–65%.
One line in it I really agree with: MCP tools give agents powerful capabilities, but every MCP tool call is, in fact, a full LLM inference round.