How do you prevent the orchestrator itself from becoming the new black box? If I’m trusting it to decide which worker decisions need my attention, isn’t a mistake in that judgment potentially worse than a worker making a bad decision?
Yeah, a mistake in that judgment can have a wider impact than a single worker mistake, and adding an orchestrator doesn’t eliminate that risk. I’d like to think that it’s still better than a human trying to parse through like a dozen sessions’ worth of noise, but it’s far form perfect.
What helps us is visibility and explicit goal tracking. Each worker is still a normal Claude or Codex session that I can attach to directly. There are per-goal checklists that the orchestrator monitors, so there’s something concrete to check progress against beyond “the AI orchestrator says it’s going well so here’s hoping.”
Our main goal is to keep the observability and control of a bunch of individual Claude sessions, while making the volume manageable. I primarily talk to the orchestrator, but I can still inspect or redirect the underlying work. It’s true that being able to inspect the workers only helps if you actually check. That said, our experience so far has made us more comfortable letting the agent run itself at this point.
What does “fully local” mean here? What leaves the machine, and does any of it pass through your servers?
Does each agent get its own worktree? and who handles getting all their changes back onto one branch?
How do you prevent the orchestrator itself from becoming the new black box? If I’m trusting it to decide which worker decisions need my attention, isn’t a mistake in that judgment potentially worse than a worker making a bad decision?
Yeah, a mistake in that judgment can have a wider impact than a single worker mistake, and adding an orchestrator doesn’t eliminate that risk. I’d like to think that it’s still better than a human trying to parse through like a dozen sessions’ worth of noise, but it’s far form perfect. What helps us is visibility and explicit goal tracking. Each worker is still a normal Claude or Codex session that I can attach to directly. There are per-goal checklists that the orchestrator monitors, so there’s something concrete to check progress against beyond “the AI orchestrator says it’s going well so here’s hoping.”
Our main goal is to keep the observability and control of a bunch of individual Claude sessions, while making the volume manageable. I primarily talk to the orchestrator, but I can still inspect or redirect the underlying work. It’s true that being able to inspect the workers only helps if you actually check. That said, our experience so far has made us more comfortable letting the agent run itself at this point.