Engineering

Agents and orchestration

I'm interested in agents that can do useful work over several steps. That means thinking about tool calls, handoffs, limits, and how to follow the run afterward.

Keeping track of a run

I want prompts to have versions and a reason for changing. When a run behaves differently, the exact instructions and tool calls should be available to compare.

Knowing when to stop

Stopping rules, budgets, and human review help keep a workflow manageable. A proposed action and an action the system has committed also need to be recorded separately.

Open questions

  • What is the exact instruction a run used, and who changed it?
  • Which tool calls are proposals, and which are committed effects?
  • What forces the loop to stop?