Most MCP demos are one tool call. Real work is usually a chain: create a project → read its ID → import media → pass the clip ID into the next operation → export → deliver.
The expensive part is often the handoff. In a normal agent loop, every result goes back to the model so it can copy an ID into the next call. Tura’s command_run Macro lets the agent describe the dependency graph once. A successful step can expose a value for a later step; the runtime resolves it only after that earlier step succeeds, while independent commands can run together.
In the ecommerce workflow from the benchmark, both configurations passed the same five checks and created the same deliverables. Tura Direct used 3 model requests instead of 11, and 56,372 total tokens instead of 262,915 (78.6% lower), despite making 11 MCP calls rather than 9. It is not saving tokens by doing less work; it is avoiding repeated model context.
For longer dependency chains, each handoff kept inside the Macro avoids another model re-entry, so the repeated-context saving can accumulate as the workflow grows. It is not a universal linear law—caching, retries, and the workflow shape still matter—but it is a useful design lever for multi-step MCP work.
The post includes the trace, benchmark limits, and the variable-inheritance rules: https://turaai.net/blog#what-we-learned-from-the-mcp-workflow-benchmark
Source code: https://github.com/Tura-AI/tura
Disclosure: I am sharing work from Tura-AI.
Top comments (0)