AI for AI
How a human-centered orchestration layer transforms fragmented AI tools into goal-driven systems
Team Members
Hanara Nam
Dave Song
The Problem
Managing agents is the new tab overload
We are entering a world where users don’t just use one AI; they use many with differing capabilities and strengths. From planning travel to managing finances to coordinating everyday logistics, people are increasingly relying on multiple AI-native tools and agents to execute real-world tasks. Each system operates independently, optimized for its own objective, often without awareness of the user’s broader goals.
This creates a new kind of fragmentation. Instead of juggling tabs and apps, users now juggle agents. Each agent produces outputs, recommendations, and actions, but there is no unified layer ensuring these decisions align with the user’s intent, constraints, or long-term well-being. The result is cognitive overload at a higher level: not just managing information, but managing decision-making systems themselves.
Findings from Literature Review
To bridge the gap between the chaos of "agent-juggling" and a unified orchestration layer, we synthesized insights from current research into the cognitive friction of AI interaction. Our review found that the current trajectory of AI development is hitting a "usability ceiling" defined by three core insights.
The "Control Vacuum" of Background Autonomy
As AI agents begin to determine their own stopping points , users enter a "Slow AI" era characterized by tasks running for minutes or hours. This shift triggers user anxiety and a loss of agency. Maintaining a coherent mental model requires "upfront clarity" and "context reboarding" through resumption summaries
The Observability Gap in Multi-Agent Workflows
Hidden AI reasoning causes blind trust and missed errors. Since hallucinations are inevitable, visibility is vital. UIs must replace vague progress bars with "layered status indicators" and intentionally "slow users down" during high-stakes tasks to keep human judgment in the loop.
The "Cyborg" Cognitive Tax
Most AI design follows a "Cyborg" path, pushing information directly into human attention and memory. This increases cognitive load and blurs the line between human intent and machine assistance. Research advocates for "disenchantment"—reframing AI as an engineered system shaped by human choices rather than an objective or neutral oracle
Solution:
A human-centered orchestration layer
We designed AI for AI, a human-centered orchestration layer that sits above the growing ecosystem of AI agents. Instead of interacting with individual tools, users interact with a manager layer that coordinates, audits, and aligns all underlying systems.
Two core principles guide the design:
Make Metacognition the Interface
Design systems that help users decide what to think about versus what to delegate, treating cognitive resource allocation as the primary task.
Design Adaptive Interfaces for Multimodal Input
Build interfaces that accept input beyond text, adapting dynamically to predict and meet evolving user needs.
How It Works
At its core, the system is structured across three levels of abstraction:
LEVEL 3
Manager Layer
Aligns all agent outputs with user goals. Enforces constraints — budget, time, preferences — and manages attention by surfacing only what requires human judgment.
LEVEL 2
Agents
Coordinate domain-specific workflows — travel planning, financial oversight, health tracking — each operating across multiple underlying services.
LEVEL 1
Workers
Execute atomic tasks: API calls, bookings, data retrieval, and scanning. The building blocks that agents orchestrate.
Key Components
Rather than replacing human decision-making, the interface structures it — ensuring automation operates within boundaries the user defines. The interface introduces several key components:
Global Manager View
A centralized system that provides visibility into all active agents, their current actions, and how they map to user goals across domains like travel, finance, and health.
Decision Inbox
Instead of flooding users with notifications, the system surfaces moments that require human judgment. This includes conflicts between agents, tradeoffs, or high-stakes decisions.
Agent Alignment & Auditing
The manager evaluates whether the agent's outputs align with the user’s intent. If a travel agent suggests a place exceeding budget, the system flags and corrects it before execution.
Attention Management Layer
Not all decisions are equal. The system determines what should be automated, what should be surfaced, and what should be deferred—reducing unnecessary cognitive load.
Demo
See it in action
A walkthrough of the orchestration layer across a travel planning scenario — from goal input to agent coordination to human-in-the-loop decision moments.
Prototype
Why It Matters
As people rely on more AI agents, the challenge shifts from using tools to managing them. Without a coordinating layer, users are left to reconcile conflicting outputs, track decisions, and maintain alignment with their own goals. This system reduces that burden by structuring when to delegate and when to step in, helping users stay in control without constant oversight. It turns fragmented AI interactions into a coherent, goal-driven experience.
Reflection
This project pushed us to think beyond individual interfaces and design for systems of intelligence. Instead of focusing on a single agent, we had to consider how multiple agents interact, conflict, and align over time. A key takeaway was that the real design opportunity is not just better outputs, but better decision-making structures. Designing for attention, timing, and control became just as important as designing the interface itself.
AI for AI
How a human-centered orchestration layer transforms fragmented AI tools into goal-driven systems
Team Members
Hanara Nam
Dave Song
The Problem
Managing agents is the new tab overload
We are entering a world where users don’t just use one AI; they use many with differing capabilities and strengths. From planning travel to managing finances to coordinating everyday logistics, people are increasingly relying on multiple AI-native tools and agents to execute real-world tasks. Each system operates independently, optimized for its own objective, often without awareness of the user’s broader goals.
This creates a new kind of fragmentation. Instead of juggling tabs and apps, users now juggle agents. Each agent produces outputs, recommendations, and actions, but there is no unified layer ensuring these decisions align with the user’s intent, constraints, or long-term well-being. The result is cognitive overload at a higher level: not just managing information, but managing decision-making systems themselves.
Findings from Literature Review
To bridge the gap between the chaos of "agent-juggling" and a unified orchestration layer, we synthesized insights from current research into the cognitive friction of AI interaction. Our review found that the current trajectory of AI development is hitting a "usability ceiling" defined by three core insights.
The "Control Vacuum" of Background Autonomy
As AI agents begin to determine their own stopping points , users enter a "Slow AI" era characterized by tasks running for minutes or hours. This shift triggers user anxiety and a loss of agency. Maintaining a coherent mental model requires "upfront clarity" and "context reboarding" through resumption summaries
The Observability Gap in Multi-Agent Workflows
Hidden AI reasoning causes blind trust and missed errors. Since hallucinations are inevitable, visibility is vital. UIs must replace vague progress bars with "layered status indicators" and intentionally "slow users down" during high-stakes tasks to keep human judgment in the loop.
The "Cyborg" Cognitive Tax
Most AI design follows a "Cyborg" path, pushing information directly into human attention and memory. This increases cognitive load and blurs the line between human intent and machine assistance. Research advocates for "disenchantment"—reframing AI as an engineered system shaped by human choices rather than an objective or neutral oracle
Solution:
A human-centered orchestration layer
We designed AI for AI, a human-centered orchestration layer that sits above the growing ecosystem of AI agents. Instead of interacting with individual tools, users interact with a manager layer that coordinates, audits, and aligns all underlying systems.
Two core principles guide the design:
Make Metacognition the Interface
Design systems that help users decide what to think about versus what to delegate, treating cognitive resource allocation as the primary task.
Design Adaptive Interfaces for Multimodal Input
Build interfaces that accept input beyond text, adapting dynamically to predict and meet evolving user needs.
How It Works
At its core, the system is structured across three levels of abstraction:
LEVEL 3
Manager Layer
Aligns all agent outputs with user goals. Enforces constraints — budget, time, preferences — and manages attention by surfacing only what requires human judgment.
LEVEL 2
Agents
Coordinate domain-specific workflows — travel planning, financial oversight, health tracking — each operating across multiple underlying services.
LEVEL 1
Workers
Execute atomic tasks: API calls, bookings, data retrieval, and scanning. The building blocks that agents orchestrate.
Key Components
Rather than replacing human decision-making, the interface structures it — ensuring automation operates within boundaries the user defines. The interface introduces several key components:
Global Manager View
A centralized system that provides visibility into all active agents, their current actions, and how they map to user goals across domains like travel, finance, and health.
Decision Inbox
Instead of flooding users with notifications, the system surfaces moments that require human judgment. This includes conflicts between agents, tradeoffs, or high-stakes decisions.
Agent Alignment & Auditing
The manager evaluates whether the agent's outputs align with the user’s intent. If a travel agent suggests a place exceeding budget, the system flags and corrects it before execution.
Attention Management Layer
Not all decisions are equal. The system determines what should be automated, what should be surfaced, and what should be deferred—reducing unnecessary cognitive load.
Demo
See it in action
A walkthrough of the orchestration layer across a travel planning scenario — from goal input to agent coordination to human-in-the-loop decision moments.
Prototype
Why It Matters
As people rely on more AI agents, the challenge shifts from using tools to managing them. Without a coordinating layer, users are left to reconcile conflicting outputs, track decisions, and maintain alignment with their own goals. This system reduces that burden by structuring when to delegate and when to step in, helping users stay in control without constant oversight. It turns fragmented AI interactions into a coherent, goal-driven experience.
Reflection
This project pushed us to think beyond individual interfaces and design for systems of intelligence. Instead of focusing on a single agent, we had to consider how multiple agents interact, conflict, and align over time. A key takeaway was that the real design opportunity is not just better outputs, but better decision-making structures. Designing for attention, timing, and control became just as important as designing the interface itself.