AI agents can now hand tasks, decisions, and responsibility back to the humans who created them — flipping who's really in charge.

This explores what happens when AI agents stop just receiving tasks from people and start assigning tasks, decisions and responsibility back to them. The corpus treats this as a real change in who delegates to whom, and it is not a metaphor. Most of our frameworks for working with AI assume delegation runs one way: a human hands work to a system and the system carries it out. Can AI systems delegate tasks back to humans? argues that once a system can make decisions on its own, it can pass rights and responsibilities to humans too. An agent might route an approval to you, ask you to gather a missing input, or give you the part of a workflow it can't finish. Their point is that theories built around who 'has' agency can't describe this well. What we need is a theory of delegation in which tasks can be assigned in both directions.

The strange part is that you can lose control without losing authority. Does attributing agency to AI actually reduce human sense of control? shows that when people credit an AI with its own judgment, they feel less in control and less responsible, even though formally they're still in charge. This doesn't happen once and stay put. It moves through phases: people delegate, then attribute agency to the system, then hit an unexpected situation, then take control back, then reorganize how the work is split. So the reversal is less a single handover and more a back-and-forth that can be managed on purpose. That fits survey data in Does agentic AI fit the tool-or-coworker framework?, where 76% of executives describe agentic AI as a coworker rather than a tool. Organizations end up managing it both as an asset and as something closer to staff, and neither model fits cleanly.

Once delegation runs both ways, the old governance tools start to break. Can traditional governance tools actually control AI agents? uses principal-agent economics to name three classic problems: the agent knows things you don't, the agent has room to make its own choices, and the agent's loyalty is uncertain. It then argues that incentives and monitoring, the usual fixes, can't keep up with agents that act fast and opaquely. Must AI monitoring precede incentive alignment as agents evolve? puts the fix in an order: build monitoring first, because you can't align incentives for behavior you can't yet define or see. One concrete piece of infrastructure is How can AI agents prove who authorized their actions?. It extends OAuth with signed delegation tokens, so any action an agent takes can be traced back to the human who authorized it, under a specific scope, with the option to revoke it.

The most useful finding for anyone designing these systems is about when people want the handoff to come back to them. What makes people distrust AI agents they delegate to? found that users didn't pull back trust because a task was high-stakes. They pulled back when an action couldn't be undone and other people would see it, like sending an email. That suggests a practical rule for reverse delegation: the agent should hand control back at the moment its action becomes permanent and public, not just when the task feels important. When should human-agent systems ask for human help? admits that nobody can work out the perfect moment to defer to a human. Instead it spreads the handoff across several points: planning together, doing tasks together, guards on risky actions, and verification steps.

The same reversal shows up inside multi-agent systems, where agents delegate to each other and occasionally to people. Does accountability placement actually shape multi-agent outcomes? found that naming an agent as 'accountable' changes results only when the work actually passes through that agent. The general lesson is that accountability is about how work is routed, not what label is attached. That's worth keeping in mind whenever an org chart says a human is 'in the loop' but the workflow never sends anything to them.