Agentic AI: Adaptive Planning is the Secret Sauce

Leveraging rich context and dynamic feedback for smarter, more adaptable AI planning.

Adaptive planning in agent systems is not just about executing a series of steps; it’s about balancing flexibility with precision while ensuring the system has the right amount of context to perform effectively. In today’s AI landscape, there’s a tendency to overestimate how much generative AI knows simply because it can produce convincing responses. However, much like an intern on day one—who has general knowledge but lacks specific insights about your organization or environment—an AI needs detailed, continuously updated context to succeed.

There are two main approaches to planning with AI agents. The interactive, step-by-step method allows the agent to execute tasks incrementally. This method is highly adaptive; after each step, the system can gather new information and adjust its subsequent actions. For example, when troubleshooting a computer’s connectivity issue, an agent might request a screenshot of your system to verify its current state, guide you through checking your Wi-Fi connection, and then ask for another screenshot. This continuous loop of action and feedback helps the agent adjust its plan dynamically to solve the problem.

In contrast, fully baked plans involve the agent generating an entire plan from start to finish before any action is taken. This approach offers consistency and efficiency, which can be beneficial in scenarios where repeatability is crucial. Consider planning a multi-day road trip where each stop is pre-arranged, such as identifying BMX tracks along the route. While this method can save time, it risks incorporating inaccuracies—like hallucinating a non-existent BMX track—if the system doesn’t have sufficient contextual data or if conditions change after the plan is made.

Context is key. When you ask an AI to perform a task, you must provide it with rich, detailed information about the environment and constraints. Take the example of connecting an old Nintendo Wii to a modern TV. Without knowing the specifics—like the absence of an HDMI port and the type of inputs available on your TV—the AI might struggle to generate an accurate plan. Similarly, if you want to know how to operate a Keurig coffee maker, a simple text prompt isn’t enough; a photo of the machine can provide critical details such as the layout of the controls and indicators. This rich contextual input transforms the agent from a generic problem-solver into a tool finely tuned to your specific needs.

Moreover, as the process unfolds, the agent must be continuously updated with new information. If conditions change—say, if someone else makes a travel reservation while your AI is planning your trip—the agent needs to be aware of these changes to avoid errors like duplicate bookings.

In my experience, the future of AI lies in finding a sweet spot between these planning approaches. A hybrid system that leverages the adaptability of interactive planning while harnessing the efficiency of fully baked plans could offer the best of both worlds. This means building systems that are not only capable of breaking down high-level goals into actionable subtasks but also of continuously receiving and processing rich, contextual updates. Only then can we ensure that AI remains both flexible in its execution and robust in its planning, truly transforming how we tackle complex tasks in our daily lives.