Agentic AI is shifting artificial intelligence from a tool to an independent partner. In this episode of Futurise, Rob Price interviews Nell Watson about her book Safer Agentic AI: Principles and Responsible Practices, exploring how to harness this shift safely.
So far traditional generative AI was about creating content such as text, images, audio, or video. In contrast, agentic AI is about taking an action. Instead of just responding to prompts, an agent can be given a task or mission. It can break the task into steps, plan, and carry it out on its own. Agentic AI builds on generative AI by adding procedures, tools, and guardrails. This helps it reason, plan and stay within limits. This shift turns AI models into true partners, like an operating system for life, rather than just creative helpers.
Nell shares an example from her own work. She uses Claude Code and the Opus 4.5 model to help design and run scientific and engineering experiments, even on quantum computers, without having to write any code herself.
The main challenge now is making agentic AI safer. Nell points out that as we let these systems work more independently, we still need real human oversight. We need a human in the loop. She and Rob mention two new risks:
- people might become too trusting and stop checking the AI’s work because it usually gets things right.
- less experienced workers might not have the skills to judge AI results, which can lead to ‘shadow AI,’ where important decisions are made by systems that no one is really watching.
To deploy agentic AI safely, it is important to set clear goals and plan for what could go wrong. Tasks should be defined in detail, including what to do in emergencies and when to stop a task. This is like how NASA manages risks in space missions. Nell notes that many failures happen because agents do not have enough context. They need to know if they are working in areas such as medicine, finance, or critical infrastructure, and what values matter most. They also need to understand if speed or thoroughness is more important, or if they should focus on trying new things or being reliable.
Her research looks at building these values into AI through constitutions and a kind of supervisory layer, like a superego, that checks the agent’s planned actions and steps in if it goes off track or misses important context. This is supported by technical guardrails and hooks. Hooks are strict rules in the system that limit what the agent can do. For example, the AI might be allowed to read emails but not send, forward or delete them. Or it could be blocked from doing anything that might erase a hard drive.
Nell explains that she finds new ways AI can fail by testing agents in ‘YOLO mode’ and then adds hooks to prevent those problems from recurring. She also points out that there is not yet a shared collection of these lessons, even though more people are working with agentic AI.
She makes a critical point: ‘safer’ AI is different from ‘safe’ AI. Total safety is not possible. Instead, we should follow the example of aviation, where every incident leads to better rules, monitoring and practices. This approach includes using AI to watch over other AI and combining several models to get a ‘wisdom of crowds’ effect.
Nell’s book addresses a frequent problem in organizations: knowing about risks but not knowing what to do next. Each chapter ends with practical action steps, and there is a website with a free tool to help diagnose AI issues and create a short report. The goal is to help organizations avoid both fear-based inaction and careless adoption, so they can unlock agentic AI’s potential while ensuring transparency, accountability, and robust safety practices drive progress forward. By embracing responsible innovation, we can shape a future where agentic AI works safely and reliably alongside us.
