Generative AI and Agentic AI

Generative AI learns statistical patterns in training data and creates new content such as text, images, code and biological sequences. Agentic AI pursues an objective across multiple steps by planning, using tools and adapting to results within bounded autonomy. An agent may use a generative model, but generation and goal-directed action are distinct capabilities.

How generative AI creates content

Generation has two stages: learning a distribution from examples and producing an output from that learned distribution.

• During training, data are encoded into units such as tokens or numerical representations. Model parameters are adjusted through repeated prediction and error correction, capturing statistical relationships in the data.
• During inference, a prompt conditions the learned probability distribution. The model repeatedly selects or samples likely elements, recombining learned patterns into an output that can be novel.
• The process varies by architecture: language models commonly predict tokens sequentially, while diffusion models iteratively remove noise to form images.
• Generated output reflects learned statistical relationships rather than an independently verified understanding of truth, so important results require evaluation.

Biological sequences as an example

Models can represent DNA, RNA or proteins as sequences and learn motifs and long-range dependencies by predicting missing or subsequent elements.

• DNA tokens may represent A, C, G and T or longer segments; RNA uses U instead of T, while proteins are represented through amino-acid sequences.
• Generation may be conditioned on prompts, known sequences or desired properties, and the model samples or selects likely subsequent tokens.
• Computational plausibility does not establish biological function, safety or viability; generated sequences require rigorous screening and experimental validation.

How agentic AI executes multi-step tasks

Agentic AI is generally a broader system combining models with planning, memory, feedback and external resources. A control loop connects reasoning to action, allowing the system to continue toward a defined objective without requiring human instructions at every step.

• The system receives an objective together with permissions, constraints and, where applicable, stopping criteria.
• A planner decomposes the objective into tasks, orders them and identifies information or resources needed for each task.
• An orchestrator maintains the current task state and asks a model or programmed component to select the next action.
• The agent translates that action into a tool call, such as querying a database, invoking an application programming interface or operating within a software environment.
• It observes the tool's result, records relevant information in its working context or memory and assesses whether progress has been made.
• If an action fails or new information changes the situation, the agent may retry, choose another tool or revise its plan.
• The loop ends when the goal or stopping condition is met, a limit is reached, or the system escalates the decision to a human.
• Limited human intervention is possible because intermediate decisions are automated, but autonomy remains bounded by permissions, available tools, programmed constraints and human oversight. Multiple specialised agents may also coordinate by assigning tasks and exchanging results.

Cyber risks, controls and accountability

• Generative AI can scale deceptive content, impersonation and insecure code. Agentic AI adds execution risk because prompt injection or manipulated data may trigger cascading actions through connected systems.
• Credentials, memory, plug-ins and external interfaces expand the attack surface, particularly when an agent can modify data, communicate externally or execute code.
• Controls include least-privilege access, tool allow-lists, isolated execution environments, input and output validation, logging and continuous monitoring.
• High-impact or irreversible actions should require human approval, while rate limits, spending limits and stopping rules constrain automated activity.
• These systems can automate defensive triage and response, but accountability remains with human organisations. Governance must therefore assign responsibility for design, deployment, supervision and incident handling.