Moving to Agentic AI

ENHANCING SECURITY AND RESILIENCE

Security and resilience depend on precision and constraint. Each agent must operate under a least-privilege model, granted access only to the tools and data required for its assigned purpose. Every interface—API calls, parameter exchanges, tool outputs—should be validated, logged, and monitored for unexpected behavior. Adversarial testing, including prompt manipulation and tool-chaining simulations, helps to identify vulnerabilities before they manifest.

Agents should not be allowed to communicate freely; information flows must be governed, with a clear separation of duties and a supervisory layer capable of detecting unauthorized coordination. When failures occur—and they inevitably will—agents must fall back to predefined safe states, with compartmentalized systems ensuring that errors do not cascade through the network.

IMPROVING EFFICIENCY AND OPERATIONAL FLOW

Efficiency and cost control arise from smart design. Organizations can reduce overhead by limiting context size, minimizing redundant API calls, and using smaller models when appropriate.

Clear termination conditions and task decomposition prevent agents from looping endlessly or over-solving problems. Reflection mechanisms and feedback loops can be used to enhance reasoning, ensuring that agents learn from experience without drifting from objectives. This balance between adaptability and constraint keeps systems effective without runaway complexity.

ENSURING AUDITABILITY AND CONTINUOUS IMPROVEMENT

Auditability remains the foundation of trust. Every agent action, decision, and permission change must be immutably logged and retrievable for review. Agents should be designed to explain their reasoning in plain language, making it clear which data and rules informed a given outcome.

Continuous monitoring systems, potentially supported by specialized “judge” models, should assess performance, fairness, and safety in real time. Regular recalibration, including retraining and memory review, prevents agents from relying on stale or biased information. This cycle of evaluation and refinement transforms AI governance from a static policy into an evolving practice. Continuous monitoring should not only measure conformance but also surface patterns that help systems and people learn from one another. Co-adaptive design treats governance as dialogue, not oversight. Agents evolve through feedback, while human operators adjust policies based on how those agents actually behave in the wild. Over time, well-designed agentic systems can become collaborative learners rather than compliance liabilities.

THE NEW CRAFT OF AGENTIC SYSTEMS

Agentic AI changes the nature of software management. It replaces static rules with adaptive systems that learn, adapt, and negotiate among themselves. Managing that complexity requires treating agents not as programmable tools, but as collaborators operating within governed ecosystems.

This shift challenges traditional IT practices yet also creates new and necessary technical roles. Those who understand both the architecture of intelligence and the art of keeping it aligned with human purpose will be in high demand.

IS IT 1995 ALL OVER AGAIN?

My 1995 forecasts were clearly aggressive: time-bound stretch goals, not inaccurate conjectures. The engineering did not meet the promise at the time.

Many of those constraints have been removed. Transformer models revolutionized knowledge representation. Our vastly more powerful infrastructures and near-infinite stores of data, however, have introduced new challenges. These challenges are less about technological capability and more about preparation, management, and the evolving relationships between humans and systems—oversight and autonomy and the interplay of design and discovery.

The vision for co-working with AI-empowered agents remains clear today. Agents embody the next step in realizing the human goal of automating the mundane. We may not yet know what we will do with the time freed by AI; however, we still desire software empowered to handle the routine, dreary, and repetitive mental work that is as numbing to the human spirit as repetitive physical work is to the human body.