Your AI agent just spawned another agent. Now who’s responsible for what happened?

The more I experiment with multi-agent coding workflows, the more uncomfortable one thing becomes: “The agent did it” is not really an audit trail. Imagine: Human → Architect → Builder → Sub-agent → Tool → Code change The builder may have been authorized to modify the repository. But did it have the authority to spawn the sub-agent? Did the sub-agent inherit that authority? Who authorized the tool call? And if something goes wrong, can you reconstruct the entire chain? Git gives you commits. GitHub gives you PRs and branch protections. But neither necessarily tells you which agent was acting under whose authority at the moment an action happened. I've been building SUTRA around this problem, and I'm increasingly thinking that agent identity, delegation and provenance may become as important as the actual agent capabilities. Not trying to solve it with another giant prompt either. I'm more interested in what belongs in the orchestration/control layer. Curious how others are approaching this: Do child agents inherit permissions? Does every agent get its own identity? Do you record delegation chains? Where do you draw the boundary between autonomy and authority? Is this actually a problem worth solving now, or are current agent systems still too small for it to matter? Would love to hear how people building multi-agent systems are handling this.