In today's column, I present a new prompting strategy aimed at optimizing the use of multi-agentic AI. As agentic AI, including generative AI and large language models (LLMs) that handle specific tasks, continues to evolve, an increasing number of these AIs will emerge. Thus, effectively composing prompts to engage the appropriate agentic AIs becomes crucial. Mismatched prompts could result in unnecessary engagements or fail to activate essential AIs. This column is part of my ongoing Forbes series, which explores the latest developments in AI and various complexities within the field. **Prompt Engineering in the Age of Agentic AI** Previously, I discussed numerous prompt engineering techniques that can enhance your results with generative AI, potentially leading to lucrative opportunities. The rise of agentic AI introduces new dimensions to prompt engineering. Consider planning a vacation with generative AI: while initial planning may be straightforward, booking often requires switching to third-party websites. Enter agentic AI, which can function as a virtual travel agent, taking care of planning and booking through natural language interactions. **The Value of Multi-Agentic AIs** Rather than relying on a single agentic AI, utilizing multiple agents can greatly enhance efficiency. However, this also raises the challenge of determining which agents to invoke. Incorrect selections may incur costs or complicate tasks, while failing to engage the right agents could hinder progress. I propose two main approaches for prompt composition involving multi-agent AI: 1. **Driver’s Seat:** Users specify which AI agents to activate and in what sequence, reducing ambiguity. 2. **Passenger’s Seat:** Users outline the overall task, allowing the generative AI to decide which agents to engage. There are trade-offs with each approach. The driver's seat provides control at the cost of potential complexity, while the passenger's seat offers simplicity but requires clear communication to avoid vagueness. **Examples in Coding Assistance** To illustrate, let's use a coding scenario where five AI agents are available: 1.

**CodeFixer:** Debugs and optimizes code. 2. **CodeReviewer:** Evaluates code for best practices. 3. **BugHunter:** Identifies vulnerabilities and logic errors. 4. **PerfAnalyzer:** Assesses performance and suggests optimizations. 5. **DocWriter:** Generates documentation. *Driver's Seat Example:* When given a prompt to fix a Python script, I might instruct: “Invoke CodeFixer, then BugHunter, followed by PerfAnalyzer. ” The generative AI then confirms and executes the specified sequence. *Passenger’s Seat Example:* Alternatively, I could say: “I need help with my Python script; please invoke agents that can assist. ” The AI interprets my request and selects the agents while keeping me informed about its planned actions. However, vague prompts can lead to less effective outcomes. For instance, simply asking for “help with my Python script” may not result in useful agent involvement unless clarified. **Emerging Research in Multi-Agent AI** Research in the AI community is advancing rapidly. A recent study introduced "AgentRec, " which focuses on efficiently selecting agents based on user prompts. This study's insights highlight the potential for generative AI to enhance its ability to choose the appropriate agents based on learned data. As we adopt and refine these techniques, practicing invoking multi-agent AIs will be crucial. While many generative AI applications currently restrict direct agent invocation, the future promises improvements. Remember Lincoln's words: "The best thing about the future is that it comes only one day at a time, " which applies to the evolving landscape of multi-agentic AI.