Agents Are Not Enough

Chirag Shah, Ryen W. White

Abstract

In the midst of the growing integration of Artificial Intelligence (AI) into various aspects of our lives, agents are experiencing a resurgence. These autonomous programs that act on behalf of humans are neither new nor exclusive to the mainstream AI movement. By exploring past incarnations of agents, we can understand what has been done previously, what worked, and more importantly, what did not pan out and why. This understanding lets us to examine what distinguishes the current focus on agents. While generative AI is appealing, this technology alone is insufficient to make new generations of agents more successful. To make the current wave of agents effective and sustainable, we envision an ecosystem that includes not only agents but also Sims, which represent user preferences and behaviors, as well as Assistants, which directly interact with the user and coordinate the execution of user tasks with the help of the agents.

Introduction: The Agency Trade-off

The AI community is currently witnessing a massive resurgence in "agentic AI"—autonomous programs designed to act on a user's behalf. However, history suggests that building smarter agents is not a guaranteed path to success. From the symbolic AI of the 1950s to the reactive agents of the 1990s, previous eras have struggled with brittleness and scalability.

The fundamental challenge lies in the "agency trade-off": the balance between the perceived benefit of an agent and the cost in terms of time, privacy, and risk. If a user must constantly intervene or clarify instructions, the agent fails to provide value. To overcome this, we must move beyond standalone bots toward a more robust, integrated ecosystem.

A Conceptual Framework for Modern Agents

Rather than introducing a single algorithm, this work proposes a conceptual shift supported by five technical pillars:

  1. Neuro-symbolic Integration: Merging the flexibility of machine learning with structured reasoning.
  2. Hybrid Architectures: Utilizing hierarchical models (combining small and large LLMs) for efficiency.
  3. Enhanced Coordination: Developing decentralized protocols for multi-agent communication.
  4. Robust Learning: Enabling agents to adapt via reinforcement and transfer learning.
  5. Ethical Design: Embedding transparency and accountability into the system's core.

The Proposed Ecosystem Architecture

The centerpiece of this framework is a three-tier architecture designed to bridge the gap between narrow tasks and personal user needs. Success lies in the synergy between these components rather than the power of a single agent.

The proposed ecosystem: A User interacts with an Assistant, which manages specialized Sims to coordinate various task-specific Agents.

1. Assistants: The User-Facing Orchestrator

The Assistant serves as the primary interface for the user. Instead of managing dozens of specialized agents, the user interacts with a single, private Assistant. This component understands the "who" and "why" behind a request, acting as a buffer that protects personal data while orchestrating complex tasks.

2. Sims: Digital Representations of the User

Unlike generic "personas," Sims are dynamic representations of a specific user. They encapsulate actual preferences, historical context, and privacy constraints. By using Sims, the system can simulate outcomes and ensure actions align with the user's intent without exposing raw personal data to external agents.

3. Agents: The Task-Specific Workers

Agents are the "worker bees" of the system. They are specialized modules designed to execute narrow, well-defined tasks (e.g., booking a flight or searching a database). They receive instructions from the Assistant, which has been informed by the user's Sim, ensuring the execution is both efficient and personalized.

Discussion and Future Outlook

While this ecosystem provides a roadmap for trustworthy AI, several hurdles remain. There is a lack of standardized networking protocols for cross-platform agent communication. Furthermore, building "Sims" that are both accurate and private is a significant technical and ethical undertaking. Trust cannot be built overnight; agents must gradually earn it through consistent accuracy and transparency.

Success in the next age of AI evolution will not come from a single, all-powerful agent, but from the synergy between these three layers. By separating intent (Sims), orchestration (Assistants), and execution (Agents), we can create agentic systems that are both highly capable and socially acceptable.

Summary updated

Abstract

In the midst of the growing integration of Artificial Intelligence (AI) into various aspects of our lives, agents are experiencing a resurgence. These autonomous programs that act on behalf of humans are neither new nor exclusive to the mainstream AI movement. By exploring past incarnations of agents, we can understand what has been done previously, what worked, and more importantly, what did not pan out and why. This understanding lets us to examine what distinguishes the current focus on agents. While generative AI is appealing, this technology alone is insufficient to make new generations of agents more successful. To make the current wave of agents effective and sustainable, we envision an ecosystem that includes not only agents but also Sims, which represent user preferences and behaviors, as well as Assistants, which directly interact with the user and coordinate the execution of user tasks with the help of the agents.
BibTeX
@article{shah2025agents,
  author = {Chirag Shah and Ryen W. White},
  title = {{Agents Are Not Enough}},
  journal = {Computer},
  year = {2025},
  doi = {10.1109/mc.2025.3575829},
}
Figure