Have you looked at whether client-side schema caching could recover that? If the agent doesn't re-transmit tool definitions on every turn, the sequential vs parallel gap should narrow significantly.
In the evaluation I did account for prompt caching, so the multi-turn penalty is already minimized, yet it was still significant enough to make the difference.
akane8•1h ago
Numbers are quite interesting. A few key findings -
1. Github MCP is 2–3x more expensive to use than CLI. Probably not surprising. There’s almost no practical reason to use their MCP except for some of the different handling of security
2. Tool Search saves upfront tokens but spends them on extra turns. Whether that trade-off pays depends on task complexity. ToolSearch also introduces a new failure mode due to imperfect search accuracy.
3. Code Mode is the cheapest way to use MCP, but still 2x more expensive than CLI, and it’s very slow. Code-mode also introduces a unique failure mode when the agent writes buggy code or poor error handling.
4. It's possible to push CLIs further towards higher success rate at lowest cost and latency with a principled design approach that treats agent ergonomics as first class concern. I detailed this in https://axi.md.
Benchmark harness, results and the reference implementation of gh-axi are all open sourced at https://github.com/kunchenguid/axi.