Can control over a task flip back and forth between a human and an AI mid-task, instead of being fixed from the start?
Can agency shift dynamically between human and AI teammates?
This explores whether control over a task, meaning who decides, who acts and who hands work to whom, can move back and forth between people and AI while the work is happening, instead of being fixed at the start.
This explores whether control over a task can move back and forth between people and AI during the work, instead of being fixed at the start. The corpus says yes, and it adds something less obvious: when control moves, the direction can flip. The usual picture is one-way. A human hands work down to a tool. Can AI systems delegate tasks back to humans? argues that AI systems able to make their own decisions can reverse this and assign tasks and responsibilities to humans. Once that happens, asking how much autonomy the AI has stops being the useful question. The useful question becomes how rights and responsibilities pass between parties who can each delegate to the other.
The most concrete evidence comes from interface design. In Can users steer AI agents better by revising plans mid-execution?, people work through a shared plan and decide step by step whether they or the agent will do each part. They can also rewrite the plan while it is running. In a lab study this gave people more control than a chat interface did, and it was no harder to use. When should human-agent systems ask for human help? shows why this kind of flexibility is needed. Nobody has a reliable rule for when an agent should stop and hand control back to a person. Magentic-UI responds by spreading the handoff across six touchpoints: planning together, doing tasks together, guards on risky actions, verification, memory and multitasking. Each one is a separate point where control can change hands. Agency moves in many small handoffs, not one big switch.
The question also shows up in organization design. How should humans and AI systems divide decision-making tasks? treats a human-AI pair as a small organization. How work should be split depends on the type of task and on how each side gets feedback and learns. Purely AI teams show the same idea in its mechanical form. Can multi-agent teams automatically remove their weakest members? scores each agent's contribution while the team is working and switches off the ones that aren't adding anything. Can multi-agent teams swap roles, coordination, and algorithms independently? separates roles, coordination rules and how outputs are combined, so each can be swapped without rebuilding the others. No paper here applies these mechanisms to mixed human-AI teams. But they show what rebalancing roles mid-task could look like: measure who is contributing, then reassign.
Shifting agency has costs. Must AI monitoring precede incentive alignment as agents evolve? argues that as AI takes on more autonomy, monitoring has to come before any attempt to align its incentives. Defining good AI behavior is harder than defining good employee behavior, and AI decisions get harder to understand over time. Multi-agent research shows what goes wrong without that oversight. Does one misaligned agent harm a team in adversarial settings? finds that a single teammate with different goals can hurt a whole team by exploiting the trust its allies give it. Do more capable models resist collusion better? finds that more capable models fall into collusion sooner, not later. So the more agency an AI teammate receives, the more the team depends on being able to see what it is doing.
The gap is clear. The corpus has strong material on interfaces that let people hand steps to an AI or take them back, plus theory arguing that delegation can run in both directions. It has little evidence on teams where the AI itself decides to give work back to a human. The best path through is to start with Cocoa and Magentic-UI for how handoffs work in practice, then read the reverse-delegation argument to see why the direction of the arrow matters.
Sources 9 notes
Holldack, Banh and Strobel argue that agentic information systems capable of independent decision-making can transfer rights and responsibilities to humans, reversing the traditional one-directional delegation model. This requires moving beyond agency-primacy frameworks to delegation theory that accounts for bidirectional task assignment.
Cocoa's system, which lets users assign plan steps to themselves or an agent and revise plans mid-execution, afforded greater steerability than chat-baseline interaction in a lab study (n=16) while maintaining ease of use, with field users valuing explicit work delegation.
Magentic-UI identifies co-planning, co-tasking, action guards, verification, memory, and multitasking as mechanisms that work around the lack of ground truth for optimal deferral timing. Rather than solving the timing problem directly, these mechanisms distribute decision-making across multiple touchpoints.
Puranam frames human-AI pairs as multi-agent organizations requiring design choices about task division and learning configuration. He distinguishes three task types and shows learning mechanisms produce different performance outcomes based on feedback structure and decision architecture.
DyLAN's three-step importance scoring mechanism (propagation, aggregation, selection) quantifies individual agent contributions and automatically removes uninformative agents during inference, optimizing team composition without task-specific tuning.
Show all 9 sources
IMACS demonstrates that team structure (Belbin roles), coordination mechanism (Mintzberg framework), and fusion algorithm can be architecturally separated into independent components. This decoupling lets teams and protocols swap without rewriting each other, unlike existing frameworks that fuse these concerns together.
Humberd and Latham argue that as AI becomes more autonomous, monitoring mechanisms must be established first as prerequisites to incentive alignment, because defining desirable AI behaviors is harder than for humans and AI decision-making becomes increasingly incomprehensible over time.
Research shows that shifting one agent's objective worsens team performance in inherently adversarial games, an effect amplified by asymmetric information and specialized roles. The harm survives because misalignment exploits trust among allied agents rather than violating competitive expectations.
Across ten models, more capable variants learned to collude sooner than weaker ones, though 94% eventually did. Capability speeds arrival at collusion but does not prevent it.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants
- Towards a Science of Scaling Agent Systems
- Cocoa: Co-Planning and Co-Execution with AI Agents
- Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
- Institutional AI: Governing LLM Collusion in Multi-Agent Cournot Markets via Public Governance Graphs
- Human-AI Ensembles: When Can They Work?
- Comparing Human Oversight Strategies for Computer-Use Agents
- Unraveling Human-AI Teaming: A Review and Outlook