What should an AI work assistant be allowed to do on its own, and what should it always check with you first?
What scope boundaries and context awareness matter most to workplace agent users?
This explores what limits on an agent's actions (what it may touch, spend, or decide alone) and what kinds of situational awareness (timing, history, workplace norms) matter most when people hand work to AI agents, though the corpus mostly answers from the design and benchmark side rather than from surveys of workplace users.
This explores which limits on an agent's actions, and which kinds of awareness of the situation, make AI agents usable at work. One caveat first: the collection has little direct research asking workplace users what they want. What it does have is a clear picture, from design proposals, benchmarks and failure analyses, of where boundaries and context break down. Read together, those suggest the two problems are closely linked.
On boundaries, the corpus is converging on one idea: users should set scope for each task, not once for the whole agent. OpenAI's Model Spec writes autonomy down as an explicit record listing the allowed tools, cost limits, banned high-risk actions and a condition for when the agent should stop. It borrows the security principle of giving software only the access it needs, and it also tries to avoid burying users in confirmation prompts How can autonomous agents act safely within defined boundaries?. Designers have reached a similar answer from the interface side. They describe an 'autonomy slider' that users set per task, with three settings: suggest only, draft, or execute Should agent autonomy be a continuous slider or fixed levels?. The shared lesson is that scope is a setting users adjust, not a fixed property of the agent. The glossary note also points out that nobody has yet tested whether a slider beats other ways of giving users control.
Context awareness turns out to be more than the agent remembering more. One note argues that AI context is mutable and short-lived. The prompt, chat history, retrieved documents and hidden state all keep shifting, so users can't learn it the way they learn a stable interface How does AI context differ from conventional software context?. That matters at work, where you need to know what the agent knows before you trust what it does. Research on proactive agents adds a social layer. Agents that are capable and adaptive but lack 'civility' (respect for timing, boundaries and the user's direction) come across as intrusive. A smart interruption at the wrong moment still feels like a scope violation How can proactive agents avoid feeling intrusive to users?. A heuristic review of deployed agents found that intent visibility, meaning whether the user can see what the agent is trying to do, is among the principles most often violated. Productivity tools did comparatively well on it Do productivity agents actually protect users better than emotional ones?.
The benchmarks show why this matters in practice. In a simulated company, leading agents finished only about 30% of tasks. The main failures were social interaction with colleagues, navigating professional software and missing domain knowledge Why do AI agents fail at workplace social interaction?. Each of these is a kind of context failure: knowing who to ask, what a screen means, and what a company's work involves. Some fixes come from outside the agent. Redesigning interfaces for agents improved task completion without hurting human usability Can interface design help agents complete computer tasks better?. A trained external 'context manager' can decide what to keep and what to cut, and stronger agents do better with fuller context while weaker ones need more compression Can an external manager handle context for frozen agents?.
The surprising thread is that scope and context may stop being one agent's problem. Several notes argue that real workplace tasks exceed what any single agent loop can organize, and point toward teams of specialized agents with supervisors overseeing them Do single agents always hit organizational limits? How should agent capabilities be organized into functional layers?. If that happens, the user's question shifts from "what can this agent do?" to "which agent in the hierarchy is allowed to do what, and who is checking?" Dubberly warns that this layering adds design risk as well as capability. The corpus does not yet show how users would understand or control boundaries spread across a system they can't see.
Sources 10 notes
OpenAI's Model Spec formalizes agent autonomy through structured scope records listing allowed tools, cost limits, and prohibited high-risk activities, mirroring software security's principle of least privilege while avoiding confirmation overload.
The AI UX Playground glossary names the autonomy slider as a continuous control letting users dial agent independence per task, collapsing agency levels into three adjustable points: suggest-only, draft, execute. The pattern shows shared vocabulary among designers, though validation against other control shapes remains absent.
AI interactions operate on a substrate of constantly shifting context—prompt, history, retrieved data, hidden state—that users cannot internalize like traditional UIs. This structural mutability demands a new design discipline centered on context engineering rather than interface design.
Intelligence and adaptivity alone create socially blind agents that interrupt poorly and override user direction. The Intelligence-Adaptivity-Civility taxonomy shows civility—respecting boundaries, timing, and autonomy—is essential to making proactivity welcome rather than intrusive.
A heuristic evaluation of nine deployed agents found that general-purpose productivity systems (ChatGPT, Claude) scored substantially higher on 14 HCI principles than emotionally sensitive or high-stakes agents (Character.ai, Consensus, Pine AI). Risk-critical principles like dependency prevention and intent visibility were most violated in companion and advisory contexts.
Show all 10 sources
TheAgentCompany benchmark shows leading agents achieve 30% task completion in a simulated workplace. Social interaction, professional UI navigation, and domain-specific knowledge are the three primary failure modes, with multi-turn task performance consistently dropping to 35% across enterprise settings.
UI-Verse benchmark testing found that augmented heuristics targeting agent perception, grounding, and control consistently improved agent task completion and efficiency. Human studies confirmed these designs preserved usability for human users with no observable regressions.
AdaCoM trains an external RL-based manager to prune and preserve context for frozen agents. The key finding: stronger agents benefit from high-fidelity preservation, while weaker agents need aggressive compression—optimal context management is agent-specific, not task-universal.
Research shows that real-world tasks requiring heterogeneous expertise, parallel execution, and independent verification exceed what any single agent loop can organize. Graph-based system abstractions are needed to distribute intelligence across specialized agents.
Dubberly proposes a two-axis stack model: a capability axis (chatbot → system prompt → RAG) adding inputs that extend what LLMs can act on, and a scale axis (monitoring agents → supervisors → orchestrator) aggregating oversight. He cautions that ubiquitous agent hierarchies raise design risks alongside capability gains.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Assistant or Actor? Student Trust, Control, and Delegation Regret When Using a General-Purpose AI Agent
- Design Principles for Human-Agent Interaction
- Levels of Autonomy for AI Agents
- Controlling AI Agent Participation in Group Conversations: A Human-Centered Approach
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- LiveMCP-101: Stress Testing and Diagnosing MCP-enabled Agents on Challenging Queries
- Cocoa: Co-Planning and Co-Execution with AI Agents
- A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments