Background
AI agent concept
An AI agent (component) is a software‑based system that perceives inputs from its environment, maintains or updates an internal state, and selects and executes actions using APIs or other tools to achieve defined goals within assigned permissions and constraints. It may plan, learn from feedback when permitted, and self‑monitor its behaviour to improve performance over time, while operating within defined governance, oversight, and control mechanisms.
AI agents operate through a cycle of structured perception, reasoning, and action. The Perceive, Reason, and Act (PRA) cycle in Figure 1, is fundamental to understanding this process.
- Perceive: captures and interprets signals or data from authorised internal and external sources.
- Reason: evaluates inputs against defined goals, constraints and policies.
- Act: action executes permissible steps through approved tools or system interfaces. Integrated feedback mechanisms enable agents to learn from outcomes, update internal state, and adapt future decisions within authorised boundaries.
Common methods and technologies for AI agent development include:
- Models: AI agents are built on a range of models, from small and large language models (SLMs and LLMs) used for text generation, reasoning, and planning, to multimodal models that jointly understand text, images, audio, and video for richer perception and action planning. Classic machine learning and deep learning models support core tasks such as classification, ranking, entity extraction, embeddings, and perception, such as neural networks and transformers. They can operate alongside task specific models for summarisation, translation, optical character recognition (OCR) and document understanding, speech recognition and synthesis, named entity recognition, sentiment analysis, and anomaly detection.
- Memory and data: Agents use augmented generation approaches, such as retrieval augmented generation (RAG) or context augmented generation (CAG), together with governed short-term and long-term memory to ground decisions in enterprise knowledge stores, leveraging vector store, enterprise smart search, and knowledge graphs.
- Tools: Agents use approved tools such as business systems, services, databases, software components, and automation services through standard interfaces, such as APIs, function calls, or web services. To keep actions safe and reliable, each action is executed in a controlled environment with appropriate access controls, such as Zero-Trust or Least-Privilege, and all tasks are observable, auditable, and can be rolled back.
- Feedback: Agents learn by trying actions and improving based on feedback. This includes reinforcement learning methods that help them make better decisions over time. Their performance can be further improved by training them on organisational specific data. Human oversight, such as people reviewing or guiding agent behaviour, also helps improve results. Feedback loops allow agents to learn safely over time, within clear rules and governance limits.
Agentic AI concept
Agentic AI (system pattern) refers to a class of co-ordinated AI agents or AI systems that include a multi-step PRA cycle that perceives and infers from their environment, maintains or updates an internal state, decides how to achieve goals through reasoning, and executes a series of actions independently to achieve predefined objectives within defined permissions and constraints. They operate with varying levels of autonomy and therefore require human oversight, safeguards, continuous evaluation, and rollback controls.
Unlike past technology innovations that focused on digitised processes, agentic systems pursue outcomes, adapt through feedback, and collaborate with humans and other agents. Agents exist on a spectrum of autonomy, often working cooperatively with other agents, and are already deployed across industries in research, coding, compliance, and customer service. For government, where structured, high-volume tasks dominate, the potential is especially significant. Agents can shift public administration from doing things right to doing the right things, delivering faster, more accountable, and outcome-driven services.
Agentic AI key components
Traditional machine learning systems and many AI agents typically operate within a single application and stop at predictions or recommendations. They usually do not plan or execute multi‑step actions across systems. By contrast, agentic AI systems add an orchestration layer that coordinates one or more agents, governs tool use, embeds human oversight, and enforces safety operations including technical guardrails, accountability, security, and governance.
The diagram provides a high‑level view of an agentic AI system in an enterprise environment. It shows how an orchestration layer routes inputs, from users or systems, into an agentic workflow comprising one or more cooperating agents.
Outputs are returned after actions are completed or queued for approval. These agents rely on governed reasoning capabilities using models, memory, data, and approved tools bounded by responsible AI technical guardrails, continuous evaluation, and improvement cycles or feedback loops. Human oversight and accountability define approval checkpoints and review-recourse mechanisms across the agentic workflow.
Foundational layers such as security and governance apply throughout, ensuring decision rights, privacy, records obligations, access controls, and auditability are upheld. The environment represents the broader context of networks, infrastructure, services, and integrations within which the agents act.
The control tower acts as central architectural governance, providing observability and supporting the operations of the AI system. Together, these components illustrate how agentic AI capabilities are coordinated, controlled, and made auditable from input to outcome.
Common methods and technologies for agentic AI
Agents are built on language and multimodal models and extend beyond single‑agent behaviour to orchestrate multi‑step, goal‑directed activity. Specialised agents collaborate under an orchestration layer, operating within responsible AI guardrails and human oversight.
Example agent types may include:
- Planner agent: Decomposes high-level goals into ordered steps, sets success criteria, and assigns work to other agents
- Executor agent: Performs permitted actions via approved tools or APIs under access controls, such as zero trust or least-privilege
- Researcher agent: Finds and summarises authoritative information from enterprise search, knowledge bases, and document stores to ground outputs using augmented-generation approach, such as RAG or CAG
- Router agent: Classifies requests and routes them to the appropriate workflow, specialised agent, or model such as small or large language models
- Supervisor agent: Monitors progress for workflow steps, pauses for required approvals, and triggers a rollback or kill switch when thresholds are breached
- Assistant agent: Interacts with users, gathers clarifications, and hands off to the planner or executor agents for further action
- Monitoring agent: Continuously observes tool failures, anomalies, costs, unsafe actions, and triggers notification or alerts
- Coordination agent: Manages task sequencing, dependency resolution, and state synchronisation across multiple agents to ensure coherent and orderly multi‑agent execution
- Communication agent: Handles controlled message exchange between agents and systems, enforcing communication protocols, context boundaries, and secure information sharing across environments.
Orchestration combines an agentic framework with a workflow engine and an event bus.
- The framework sets the roles, tools, and memory. Examples include LangChain, LangGraph, CrewAI, AutoGen, LlamaIndex, or Haystack
- The workflow engine adds state, retries, and approvals. Examples include Temporal or Apache Airflow
- The bus decouples components, keeping multi‑agent workflows bounded, governed, auditable, recoverable, and portable. An example includes Apache Kafka.
Human oversight preserves accountability, legal review rights, and public trust in agentic AI systems; governance should calibrate how much oversight is required and when authorised humans intervene.
Oversight modes include:
- human‑in‑the‑loop (HITL) for pre‑approvals on sensitive actions, such as payments, entitlements, or mass notifications
- human‑on‑the‑loop (HOTL) for real‑time supervision with the ability to pause or override
- human‑out-of-the‑loop (HOOTL) for periodic audits and post‑action sampling.