Skip to content

Agent Architecture

An agent is an LLM combined with a perceive-reason-act loop that enables it to take actions in the world, observe results, and iterate. The ReAct framework (Yao et al., 2023) formalises this as a thought-action-observation cycle: the model reasons about the current state, selects an action (often a tool call), observes the result, and continues.

Memory Types

Agents require memory to function beyond single-turn interactions. Four types of memory serve different purposes:

In-context memory: The current context window contents. Immediate but size-limited and volatile.

Episodic memory: Stored records of past interactions, retrievable by relevance. Enables continuity across sessions but requires retrieval infrastructure.

Semantic memory: Domain knowledge accessed via retrieval-augmented-generation or knowledge-graphs. Provides factual grounding independent of interaction history.

Procedural memory: Reusable skills, templates, and patterns the agent can invoke. Encodes how-to knowledge rather than what-is knowledge.

Collaborative Memory

For nuclear operations, a fifth category emerges: collaborative memory where human operators and AI agents jointly build and maintain shared knowledge stores. Operators can annotate, correct, and validate agent-stored information. This combines human domain expertise with AI's ability to systematically organise and retrieve large volumes of data.

Temporal memory organisation structures information by operational periods — shifts, days, outage phases — matching how nuclear operations naturally segment time.

Autonomy Scale

Agent architectures exist on a spectrum of increasing autonomy:

  1. Chatbot: Responds to queries, no actions
  2. Assistant: Responds with access to tools, human-initiated
  3. Tool-using agent: Autonomously selects and chains tools
  4. Persistent agent: Maintains state across sessions, background monitoring
  5. Autonomous agent: Self-directed goal pursuit with minimal human involvement

Key transitions on this scale: the introduction of tool-calling (from assistant to agent), the addition of persistence and heartbeat mechanisms (from reactive to proactive), and the shift from human-initiated to agent-initiated action.

For nuclear applications, the graded-autonomy-tiers framework maps these architectural capabilities to regulatory requirements, with higher autonomy levels requiring proportionally more rigorous safety justification and human-authority controls.