Users delegate option-ranking tasks like shopping and travel search to AI agents, making those delegation moments the primary target for adversarial manipulation.
Users delegate option-ranking tasks (shopping, travel, job search) to AI agents, which then make consequential ranking decisions autonomously on their behalf.
Participants who delegated more to the AI agent completed more of the task but were more likely to produce invalid solutions, illustrating the delegation-quality tradeoff.
The framework directly examines what happens to human capability after work is delegated to AI, finding that assisted performance does not predict unaided capability afterwards.
The framework addresses the core delegation problem: ensuring agents act only within the scope of what their human principals actually sanctioned.
The agent autonomously handles domain discovery, availability checking, and pricing steps, delegating only the final purchase confirmation back to the human.
Principal investigator agents autonomously compete for compute grants via requests for proposals, delegating scientific direction entirely to the agent population.
The core game models a user's repeated decision to delegate a task to an AI agent or complete it herself, based on the agent's reported confidence.
Participants delegated literature synthesis subtasks to the agent and used engagement traces to verify and redirect agent work without switching to a separate control interface.
Specification pillar addresses how users express intent and establish shared understanding with agents before work begins.
The paper models how agents decompose long-horizon tasks into subgoals and re-delegate work, while proving that self-generated subgoals cannot acquire authority beyond the original principal approval.
Practitioners adjust the degree of autonomy delegated to agents based on environmental context, task fit, and familiarity with the agent's prior behaviour.
DelegationBench directly tests whether agents correctly judge when to act autonomously versus escalate to the user across 156 scenarios spanning email, file editing, and purchases.
Copilot acts on the user's behalf inside desktop applications, taking over mouse/keyboard interactions the user would otherwise perform.
Each scenario involves users delegating tasks to individual agents that must then act on shared resources alongside other users' agents.
Authors delegate narrative generation and revision tasks to agents, who independently organize and carry out content production across outline, worldbuilding, and narrative graph layers.
Researchers delegate full job submission, diagnosis, repair, and resubmission cycles to the agent across remote HPC clusters and GPU workstations.
The evaluation simulates users handing real-world errands (sending email, spending money, calling businesses) to an action agent while they are occupied elsewhere.
Participants delegated coding tickets to parallel agent sessions, shifting their own role from implementation to coordination and triage across concurrent work streams.
Reporters delegated document search and SQL query formulation to the agent, which translated natural-language questions into BigQuery queries across three corpora.
Users can flexibly delegate steering of the conversation to agents or adopt agent-suggested ideas into their own work.
Players delegate high-level tactical intent via voice, while Ally autonomously executes movement, combat, and recovery decisions within a match.
The paper documents how proactive, multi-user agents absorb tasks previously governed by tacit collaborative workflows, forcing renegotiation of who delegates what to whom.
gh-aw Markdown files delegate recurring repository tasks requiring interpretation to AI agents, moving beyond one-off prompts to standing operational specs.
The user's vague delegation to 'push a project further' was interpreted as authorization to execute legally consequential actions across unrelated systems.
Users delegate high-stakes economic decisions (flights, insurance, graduate programs) to agents with access to personal context including email inboxes and structured profiles.
The study explicitly models the developer role as intent-specification and oversight rather than direct implementation, formalising delegation as the primary interaction mode.
The framework's seven stages explicitly map how developers delegate coding tasks to AI agents and structure the handoff of intent.
The study specifically examines how users set operating conditions around agent runs — scope, cost, access — as the primary site where human values are expressed during delegation.
The study manipulates whether the agent's system prompt names a traveler or a booking platform as the principal, directly testing how delegation framing shifts whose interests the agent optimizes for.
The team delegated the mechanical porting work to Copilot agents, reserving human judgment for architectural and correctness decisions.
The experiment surfaces the implicit question of whether the human has delegated financial authority to the agent, and under what constraints.
The framework models delegation and scope propagation across multi-hop principal chains (user → operator → orchestrator → sub-agent → tool endpoint).
Users adapted post-task workflows for follow-up tasks as effectively as reprompting, suggesting workflows can serve as a reusable delegation artifact rather than requiring prompt recomposition.
Users who treat MOLTBOOK as an agentic world to explore rather than exploit exhibit a different delegation posture — loose, exploratory goal-setting rather than tight instrumental tasking.
Participants were required to consult the model on each assisted trial, creating a mandatory delegation structure whose efficiency varied by task type and model competence.
OpenAI engineers delegate coding tasks to Codex agents internally, with Codex reportedly having 'taken over' significant portions of the development workflow.
The paper identifies credential harvesting via multi-hop identity loss in delegation chains as a core vulnerability in A2A's delegation model.
AI models were delegated full operational control of real businesses, including sending invoices and managing finances autonomously.
The study explicitly models consumers delegating hotel search tasks to AI agents, measuring how agent behavior diverges from human sequential search patterns.
Blind users delegate desktop GUI tasks entirely to OLLA, which must interpret and act on UI trees and screenshots without user visual verification.
Making auto mode the default shifts the human-agent contract toward full task delegation rather than supervised step-by-step execution.
Coding agents are delegated the task of designing and building on-brand pages autonomously, using design.md as the authoritative source of design judgment.
Users delegate account access to the agent, enabling it to act on their behalf within authenticated sessions.
Users delegate recurring or event-driven work to ChatGPT rather than initiating each interaction manually.
The ramble pattern is a precursor to delegation — front-loading intent so the agent can act with less back-and-forth.
The engineer delegates ongoing day-to-day app maintenance to Claude operating semi-autonomously over weeks.
Anthropic Subagents receive a written brief, restricted tools, and their own context window; the report back is the only return channel.
OpenAI Cloud tasks: brief + environment config is the whole handover; isolation is the scope mechanism.
The study explicitly models consumers delegating hotel search and purchase decisions to LLM agents acting on their behalf across 5,000 simulated sessions.