On “Agents are Not Enough” — A Practitioner’s Perspective!!

 

arXiv:2412.16241v1 [cs.AI] 19 Dec 2024

I am pretty involved in working with and developing agents and, lately, have a feeling that something is missing; I chanced upon this paper by Chirag and Ryen. I took a printout and embarked on my journey to Luxor. As I was reading it, I was making notes on the margins. Here are my consolidated notes.

Cognitive Architectures

The research paper discusses the limitations and challenges of current cognitive architectures, such as SOAR and ACT-R, which were designed to model human cognition by integrating perception, memory, and reasoning. These are sophisticated designs, but these architectures have struggled with scalability and real-time performance, often resulting in high computational costs and limited practical applications.

Margin of page 2 ☝️

Figure 1 — Envisioning a new eco-system with Agents, Sims, and Assistants from page 3.

Figure 1: Envisioning a new eco-system with Agents, Sims, and Assistants.

The paper notes that while some simple rule-based agents like Alexa and Siri are widely used, and multi-agent frameworks like Swarm Robotics and AutoGen have had various successes, there is a lack of agentic systems that can score high on both capabilities (solving complex tasks) and applicability (wide range of scenarios, modalities, and contexts). The authors attribute this to five key reasons:

1. Lack of generalization: Many AI agents are designed for specific tasks and fail to generalize across different domains, due to their reliance on predefined rules and lack of adaptive learning mechanisms.
2. Scalability issues: As the complexity of tasks increases, the computational resources required by AI agents grow exponentially, hampering their ability to handle real-world applications effectively.
3. Coordination and communication: In multi-agent systems, ensuring seamless interaction and communication among agents remains a significant challenge, often leading to inefficiencies and conflicts. Additionally, there is a need for enhanced mechanisms between users and agents to ensure appropriate and effective recommendations.
4. Robustness: Many AI agents are brittle, performing well under specific conditions but failing when faced with unexpected situations, due to their lack of robust learning and adaptation capabilities.
5. Ethical concerns and safety: Ensuring that AI agents operate ethically and safely is a major concern, as failures in this area can lead to unintended consequences, such as biased decision-making or harmful actions. The trade-offs between agent control and user agency/learning opportunities are not well understood.

The paper highlights the need for advancements in cognitive architectures and AI systems to address these limitations and challenges in order to develop agentic systems that can effectively solve complex tasks while being widely applicable across diverse scenarios, modalities, and contexts.

Margin of pages 1 and 2 ☝️

CAN WE FIX AGENTS?

The research paper discusses the shortcomings of agents and agentic AI, and proposes five key directions to address these issues.

Firstly, the paper suggests integrating machine learning and symbolic AI to enhance the adaptability and reasoning capabilities of AI agents. Machine learning can provide the flexibility to learn from data, while symbolic AI can offer structured reasoning and explainability.

Secondly, the paper recommends implementing new architectures, such as caching solutions that store and execute agent workflows to reduce the need for calls to foundation models, and hybrid and hierarchical architectures that integrate small and large language models. These approaches can improve the scalability and efficiency of AI systems by decomposing tasks into sub-tasks and assigning them to specialized agents.

Thirdly, the paper proposes developing advanced coordination mechanisms, such as decentralized control and negotiation protocols, to improve the performance of multi-agent systems. These mechanisms can facilitate better communication and collaboration among agents.

Fourthly, the paper suggests incorporating robust learning algorithms, such as reinforcement learning and transfer learning, to enhance the adaptability of AI agents. These algorithms enable agents to learn from their experiences and apply knowledge across different tasks.

Finally, the paper emphasizes the importance of ethical and responsible AI design. This involves implementing guidelines and frameworks that prioritize transparency, fairness, and accountability. Integrating explainability into system design and robust testing with system deployment can help mitigate ethical and safety concerns.

The paper presents a comprehensive set of strategies to address the shortcomings of agents and agentic AI, with the goal of improving their performance, adaptability, and responsible development.

Margin of pages 3 ☝️

WHY AGENTS ARE NOT ENOUGH

The paper highlights that while addressing the technical challenges is crucial, it is not sufficient to enable the widespread adoption of capable agents. It outlines five key aspects that need to be addressed for agents to be successful:

1. Value generation: Agents must provide enough perceived benefits to users to justify the costs and risks, such as the need for frequent user intervention or privacy trade-offs.
2. Adaptable personalization: Agents must be able to adapt to the unique needs and contexts of each user, handling subtasks autonomously or seeking user input as appropriate.
3. Trustworthiness: As agents become more capable, users will need to develop strong trust in them, which will require increased accuracy, transparency, and gradual familiarity over time. Current hesitance around automatically generated emails highlights the challenge of establishing this trust for more critical agent-based tasks.
4. Social acceptability: For agents to operate at scale across diverse populations and cultures, there needs to be wide social acceptance of agent-based interactions and transactions, which may take a long time to materialize, as seen with the adoption of online bill payments.
5. Standardization: The decentralized development and deployment of agents will require efforts to standardize how they are deployed, connected, and served, similar to the development of networking protocols or app stores, to ensure compatibility, reliability, and security.

The paper emphasizes that while technical challenges are not trivial, addressing these broader issues beyond just the technology is crucial for the successful rise of capable and widespread agents.

Margin of pages 1, 2, and 3 ☝️

A NEW ECOSYSTEM WITH AGENTS

The proposed research paper introduces a new ecosystem built around agents, sims, and assistants to address the challenges of agentic AI. The key components of this ecosystem are:

1. Agents: These are narrow, purpose-driven modules trained to perform specific tasks. Agents can operate autonomously but have the ability to interface with other agents.
2. Sims: Sims are representations of the user, created using a combination of user profile, preferences, and behaviors. Sims capture different aspects of the user and can have varying privacy and personalization settings. Sims can act on the user’s behalf to accomplish tasks, coordinated by the user’s Assistant.
3. Assistants: Assistants are programs that directly interact with the user, have a deep understanding of the user, and can call Sims and Agents as needed to reactively or proactively accomplish tasks and subtasks for the user. Assistants are a private version of an agent, tailored to the user’s personal information and preferences.

The interaction between Agents, Sims, and Assistants is characterized by a high degree of synergy. The Assistant co-creates and manages Sims with the user’s supervision, reflecting the user’s multifaceted life. These Sims then engage with specialized Agents to perform tasks efficiently. This layered approach ensures that tasks are handled with precision and personalization, enhancing overall user satisfaction.

The proposed ecosystem aims to address the challenges of agentic AI, such as user information protection, user representation for autonomous task completion, and the ability for agents to communicate and negotiate on the user’s behalf to accomplish complex tasks without added burden on the user.

Margin of pages 3, and 4 ☝️

FUTURE OF AGENTIC AI

The research paper discusses the future of agentic AI, which the authors believe represents the next stage in the evolution of capable AI systems. However, the authors argue that simply building more capable agents is not enough to ensure their widespread applicability and acceptance.

The paper identifies the need to develop agents that go beyond information retrieval and generation, and can engage in reasoning and take actions on a user’s behalf. Additionally, the authors highlight the importance of developing mechanisms to enable more meaningful interactions among agents, as well as between users and agents, which they refer to as “Assistants.”

The authors also emphasize the need to address personalization, privacy, user agency, value generation, and trustworthiness of agents. To address these issues, the paper envisions a new ecosystem that includes different types of agents, as well as constructs such as “Sims” and “Assistants.”

The authors suggest that, similar to an app store, there could be an “agent store” with vetted agents available for users or their Assistants to interact with and accomplish various tasks. While agents may be the centerpiece of this ecosystem, the authors believe that the availability of Sims, Assistants, and the set of protocols that connect them will be crucial for the successful evolution of agentic AI.

The paper argues that the future of agentic AI goes beyond simply building more capable agents, and requires the development of a comprehensive ecosystem that addresses the various challenges and complexities involved in integrating these agents into our lives in a meaningful and trustworthy manner.

Margin of pages 3, and 4 ☝️

This paper is very timely, and I am working on a stealth project with a start-up on simulacrum or digital beings—more details to come.

Comments are welcome!