A Framework for Navigating the AI Agent Landscape: From Task Executors to Strategists

The views expressed in this blog post are solely my own. 

In the last few quarters there has been a lot of excitement around AI agents. Given a task or strategy by a human user, AI agents have the goal of autonomously making decisions and taking actions on your behalf. AI agents bear the promise of significant business impact and improved user experience, for example providing instantaneous, personalized 24/7 customer support. 

The idea of agents that interact with an environment has been around for many years in the area of reinforcement learning. However, with the progress and widespread adoption of LLMs, agent-based systems have seen a resurgence of interest as LLMs can provide a highly accessible and user-friendly interface for humans to interact with agent-based systems.

Many flavors of AI agents are being envisioned currently, which greatly differ in technical feasibility and level of AI autonomy. To bring more structure and clarity to the ongoing conversations on AI agents, I built a framework for navigating the AI agent landscape that I am sharing below. 

AI agents: a world of diverse applications

AI agents have the potential of wide-ranging use-cases. Here are just a few examples: 

• When it comes to customer support, recent demos have shown how AI agents can be used to provide personalized customer support 24/7 without wait times.

• More advanced AI agents have been proposed that can act as a chief of staff, who learns your preferences and how you make decisions and then seamlessly interacts with coworkers to complete tasks and makes decisions on your behalf, similar to a human chief of staff. 

• And last year, Sam Altman predicted that we will see 10-person billion dollar valuation companies soon. Imagine an entire department consisting of one human and hundreds or even thousands of agents forming a multi-agent network, all working together towards a joint objective.  

Needless to say, for these AI agent-based systems to work in the real-world and at scale, they will also require a high level of trust, accuracy and guardrails. 

A framework for navigating the AI landscape

If we reflect on the above mentioned examples it becomes clear that AI agents can have different levels of autonomy. To capture these important distinctions, which have direct impact on feasibility and enterprise level rollout strategies, I built a framework (Figure 1) for AI agents, with agent autonomy increasing from the bottom to the top. 

• Task Executor: Given the inputs of a task and a predefined process, the agent executes against it. 

• Process Selector: The AI agent autonomously chooses and executes against multiple predetermined processes. 

• Solution Finder: The input is a high-level strategy. The agent autonomously identifies possible solution paths in a constrained environment. Here, no predetermined processes are provided to the AI agent. 

• Strategist: Given high-level objectives, the AI agent autonomously identifies best high-level strategy, and subsequently identifies and executes on the best solution path.

Most AI agents launched at scale today are task executors or process selectors. Customer support is a prime example where AI agents thrive with current technology, as the decision paths for human representatives are heavily scripted and documented. Similarly, AI agents are being developed for common enterprise workflows, such as the standard software engineering lifecycle.

There is a significant technological leap, however, from a process selector to a solution finder, which requires the AI agent to dynamically observe and learn relevant contextual abstractions from its environment. As a result, most AI agents launched at scale today are task executors or process selectors. The ability to generalize from observations will be crucial for agents to move beyond predefined or common processes and workflows to genuinely solve novel or highly contextual problems and drive impact at scale.

Key ingredients for enterprise AI agents

Now that we have covered the foundations for AI agents, let’s look at some key ingredients that are necessary to build an AI agent at scale in an enterprise setting. 

Security and access controls 

What tools and data is the AI agent allowed to have access to? What information is a particular human user allowed to see? For example, if this is an AI agent who functions as a chief of staff to the CEO, it may have access to highly sensitive information. As this AI agent interacts with other employees in the organizations or other employees’ AI agents, the AI chief of staff should not leak sensitive or secret company information. 

Scalability

Hand-configuring and hand-coding each agent in a multi-thousand agent network is not scalable. Similarly, expecting data and documentation across a large enterprise to be pristine and updated in real-time for AI agent consumption is unrealistic. For AI agent systems to work at scale in complex, real-world ecosystems, they will need to dynamically learn from their environment, much like a new employee onboarding into a large organization. This includes capabilities such as participating in video calls and automatically classifying information sensitivity for proper distribution.

Human-AI collaboration

Current AI systems are imperfect, often exhibiting hallucinations, typically lacking causality, and missing contextual nuances. A human-AI team approach harnesses the complementary strengths of both. This expert-in-the-loop model enables real-world AI deployment at scale, allowing for immediate value even as the AI continues to improve.

Trust, transparency and safety

For humans to feel comfortable adapting this technology, they need to understand what is happening and be able to provide feedback to course correct the AI systematically. And as agents move into higher levels of autonomy, alignment will become evermore important to ensure that AIs act in the best interest of humanity.  

Conclusion

While AI agents have garnered lots of visibility, launching AI agents in a complex, real-world enterprise setting requires significant consideration and in some instances, the underlying AI technology is yet to be developed.