Voice agent is the orchestration layer with:

Tools:

• spawn: Able to spawn subagents that are fired off to work on particular coding tasks.

  • Params:

    • prompt

    • identifier

      • Identifier should be a natural key that is easily identifiable given the context of our check prompt

    • number (optional)

      • Should support spawning many sub agents for a single task

  • Example: “build me a website about …”

• check: manual peak into the state of the subagent.

  • Params:

    • agent identifier

  • Example: “how is our web builder agent doing?”

Skills:

• GitHub CLI

Open questions:

• How does the orchestration layer know when a subagent is finished with its task?

• What if multiple subagents finish at the same time?

  • Do we prioritize based on a certain criteria?

  • Queue them one by one?

• How do we manage our voice agent’s context relative to the spawned tasks?

  • Ideally we do not need to start a new voice session for each task we are working on.

Thoughts:

• I like the idea of the voice agent being my interface into the computer. It should be generic and open-ended.