The problem
A typical MCP coding-agent loop is often: inspect → patch → build → test → lint.
Even when the work is predictable, that means the model wakes up after every tool result to decide the next call. In a simple five-step workflow, that is 5 LLM turns.
What Tura changes
Tura exposes a macro tool called command_run. The agent sends a dependency-aware execution plan once, and the runtime executes the related commands in order. Build, test, and lint still run; the difference is that the model does not need to re-enter between each predictable step.
For the illustrated task, the loop goes from 5 model turns to 1 — an 80% reduction in model re-entries. This is a reduction in conversational/tool-calling overhead, not an attempt to skip verification.
Disclosure: I work on Tura.
- Task benchmark: https://turaai.net/benchmark-task?task=workflow-ecommerce-ad-package#runs
- GitHub: https://github.com/Tura-AI/tura
Curious whether others are tackling repeated tool-loop overhead in MCP agents this way.
Top comments (0)