
After this lesson you can name the seven components of an AI agent and what each does; use CoALA's distinction between internal (memory) actions and grounding (external) actions; map each component to its typical security weakness and OWASP category; and describe how an agent should be given an identity, an owner and an inventory entry in a shipping company.
Lesson 15 showed the agent loop from the outside. This lesson opens the box.
We take the agent apart into seven components, using a research framework called CoALA, and ask of each one: what does it do, and how can it go wrong?
1. Why This Matters
Lesson 15 described what an agent does. To secure it, we need to know what it is made of — because each part fails in its own way and needs its own control. "The AI did it" is not an answer an incident report can accept; "the tool layer allowed a write the agent should never have had" is.
2. What Is It? — A Plain Definition
In 2023 Sumers, Yao, Narasimhan and Griffiths proposed CoALA (Cognitive Architectures for Language Agents), a framework that describes a language agent through "modular memory components, a structured action space to interact with internal memory and external environments" and "a generalized decision-making process to choose actions". [CoALA] This lesson uses a simpler seven-part view built on the same idea.
A ship analogy: the bridge team. The officer decides (model); the chart table and the conning display show what matters now (context); the logbook and standing orders hold what was learned before (memory); radar, ECDIS and the radio are the instruments (tools); the passage plan sets the route (planning); the helm order carries it out (execution); and every position fix checks the result (evaluation).
3. How Does It Work? — Seven Components
CoALA also separates internal actions on memory — retrieval (reading long-term memory into working memory), reasoning (using the LLM to update working memory) and learning (writing to long-term memory) — from grounding actions on the outside world. [CoALA] For security this split is useful: internal actions change what the agent believes; grounding actions change the world.
4. How It Evolved
Early LLM applications were a single prompt and a single answer. ReAct (2022) added a loop of reasoning and actions [ReAct]; frameworks then added memory and tool layers; CoALA (2023–2024) gave researchers a shared vocabulary for the parts. [CoALA]
Engineering guidance moved in the same direction. Anthropic's three principles for agents are simplicity in design, transparency by showing the agent's planning steps, and careful design of the agent-computer interface through thorough tool documentation and testing. [Anthropic] Later guidance added context engineering: compaction, note-taking outside the window and sub-agents with clean context. [Anthropic]
5. Real Security / Engineering Context
Mapping each component to its typical weakness turns a vague "AI risk" into a list an engineer can work through. The OWASP categories in brackets come from the Top 10 for Agentic Applications and the LLM Top 10. [OWASP ASI]
OWASP's LLM06 controls apply across the tool layer: minimise extensions and their functions, avoid open-ended tools such as unrestricted shell or URL access, run with the user's own least-privilege access, and enforce authorisation in the downstream system ("complete mediation") rather than relying on the model's judgement. [OWASP LLM06]
6. Current State
Most production agents today are assembled from these parts by the application developer: a hosted model, a tool list, a context-management strategy and, increasingly, a memory store. Vendors ship building blocks — for example Anthropic documents a client-side memory tool in which the model only requests file operations and the developer's application carries them out on storage it controls. [Anthropic docs]
That division of labour matters: whatever the model asks for, it is the application around it that executes — and the application is where security controls can actually be enforced.
7. Technical Interpretation
Give every agent its own identity — a service account with its own credentials, not a person's token and never a shared admin account. Then every action in every downstream log can be traced to that agent.
Assign an owner per component: the model and prompt to the AI team, tools and their permissions to the system owners, memory to the data owner, and evaluation to whoever signs off the agent's output.
If the planning steps are not logged, the agent cannot be investigated. Transparency is not only a design principle; it is the audit trail.
8. Maritime Connection
A ship already works this way: decisions, instruments, records and checks belong to defined people under the Safety Management System. An agent should be described the same way before it is allowed near fleet data.
The asset-inventory point is the author's reading of UR E26, which does not mention AI agents.
Why what goes into the model's window matters as much as the model itself.
9. Security Considerations
An agent acting with a person's full token inherits all of that person's rights (OWASP ASI03). Use a dedicated, least-privilege identity. [OWASP ASI]
Treat every tool result as data that may contain instructions; never let it change the agent's goal or permissions.
If the agent can write and run code, run it in a sandbox with no network or credentials by default (OWASP ASI05). [OWASP ASI]
Authorisation must be checked by the target system on every call, not assumed because the agent 'decided' it was fine. [OWASP LLM06]
- An agent = model + context + memory + tools + planning + execution + evaluation.
- CoALA separates internal actions (retrieval, reasoning, learning on memory) from grounding actions on the outside world.
- Anthropic's principles: simplicity, transparency of planning steps, and careful tool (agent-computer interface) design.
- Each component has its own attack surface — goal hijack, poisoned context or memory, tool misuse, unchecked code, abused identity.
- Give each agent its own least-privilege identity, an owner per component, and an inventory entry; enforce authorisation in the target system.
Lesson 17 — Multi-Agent Collaboration puts several agents together: planner, research, analysis, security and reporting agents — what they gain, what they cost and how failures spread between them.
The Cybersecurity side of ShipPaulJobs continues in parallel: see the Maritime Cybersecurity Learning Roadmap.
Sources checked November 2026. Vendor statements are the vendor's own claims.
- Sumers, T. R., Yao, S., Narasimhan, K., Griffiths, T. L. "Cognitive Architectures for Language Agents", arXiv:2309.02427 (TMLR version, 2024) — arxiv.org.
- Yao, S. et al. "ReAct: Synergizing Reasoning and Acting in Language Models", arXiv:2210.03629 (ICLR camera-ready version, 2023) — arxiv.org.
- OWASP GenAI Security Project (9 Dec 2025). OWASP Top 10 for Agentic Applications for 2026 — genai.owasp.org.
- OWASP GenAI Security Project. LLM06:2025 Excessive Agency — genai.owasp.org.
- IACS UR E26 Rev.1 (Nov 2023), Cyber resilience of ships. Sections 4.1.1 and 4.2.6 — IACS PDF.
- Anthropic (19 Dec 2024). "Building effective agents" — anthropic.com/engineering.
- Anthropic (29 Sep 2025). "Effective context engineering for AI agents" — anthropic.com/engineering.
- Anthropic. Memory tool — Claude API documentation — platform.claude.com/docs.
Maritime cybersecurity consultant at ShipPaulJobs, writing on IACS UR E26/E27, maritime IT/OT security and how AI changes security operations at sea and ashore.
⚓ Join the ShipPaulJobs Community
Join →
Comments
Post a Comment