Selection is part of the system
My router chooses a path for a task by cost, remaining quota and previous failures. It is not my own language model. I design and operate the decision layer around existing models and coding tools.
Rules before execution
Hooks enforce routing and safety rules. The broker caps paid calls with a daily budget. A model decision should not be able to override these boundaries. The architecture needs to make it clear which component selects and which executes.
Handoffs across tools
My own MCP server provides shared tools, including a mailbox between Claude, Codex, Pi and Antigravity. This lets one agent hand its state to another. The shared task still needs clear ownership.
Hand over routine work, weigh decisions
I want Harness to make routine decisions on its own. For more consequential questions, I have a council of models consider different options. That helps me weigh them, while responsibility for the outcome stays with me.