Safer Agentic AI: Read an excerpt from Nell Watson’s book on guardrails

As we outsource more of our lives, and our world, to artificial intelligence, what safety mechanisms will we need to put in place? An excerpt.

Agentic AI development presents a major governance challenge requiring a sophisticated and coordinated response. Advancements in AI alignment, scalable oversight, and reward modelling will be essential. Systems must be designed to understand and act according to human preferences, even in ambiguous situations. Ordinary users will presumably be tasked with defining and enforcing constraints on AI behaviour — a potentially immense responsibility.

.

The path forward is full of potential but fraught with risks. Technical research into AI safety is indispensable. We need governance frameworks that can keep pace with rapid technological advancement. Ethical guidelines must be established and rigorously followed. Public engagement is necessary to ensure informed societal decision-making. Addressing these challenges requires a concerted effort from researchers, policymakers, industry leaders, and the public alike.

Success will depend on a multi-stakeholder approach: implementing technical safeguards, establishing clear regulatory frameworks, maintaining ongoing public dialogue, and fostering international cooperation to prevent harmful arms races in AI development. These elements must work in concert to create a robust framework for responsible development.

Agentic AI marks a profound shift in the nature of artificial intelligence, bridging the gap between narrowly specialised systems and hypothetical, fully general intelligences. By enabling autonomous goal pursuit and adaptive planning, these systems promise new levels of efficiency — yet they also carry significant risks, from misaligned objectives to potential societal disruption. Harnessing agentic AI safely demands thoughtful governance, rigorous oversight, and a commitment to aligning these powerful technologies with human values. Agency demands accountability, no matter the substrate in which it lies.

To make this tangible, we propose a practical roadmap:

1. Establish clear alignment goals. Before deployment in high-stakes domains, define explicit objectives, constraints, and ethical boundaries for the agentic system. Incorporate feedback from domain experts, affected users, and frontline practitioners to ensure that goals evolve with real-world complexity.

2. Prioritise scalable oversight. As autonomy increases, so too must our ability to monitor and intervene. Continuous auditing, explainability mechanisms, and emergency override channels should be engineered into the system’s fabric.

3. Mitigate common-mode failures. Many AI systems are trained on overlapping datasets and may therefore inherit the same flaws. Diversity in training sources, testing environments, and design paradigms is crucial to prevent catastrophic synchronised errors.

4. Integrate ethics and security from the start. Responsible AI cannot be bolted on after the fact. Security hardening, bias mitigation, and ethical compliance should be core design pillars, verified at every iteration.

5. Cultivate public trust. Transparent communication about risks, limitations, and corrective actions will be vital for societal acceptance. Public-facing reporting of safety incidents, akin to aviation’s open-data culture, would advance collective learning.

Agentic AI offers humanity extraordinary leverage — if, and only if, we master the discipline of aligning agency with accountability. Governance, in this sense, becomes not bureaucracy but civilisation’s immune system.

(Excerpted with permission from Safer Agentic AI by Nell Watson and Ali Hessami, to be published by Kogan Page; 2026)