Agent architecture
AI agents need memory, boundaries and review
Calling a system an agent does not solve its architecture. It must know what to remember, what it may do, when to stop and who can review its path.
01
An agent is a loop, not a personality
In practice, an agent observes a state, chooses an action, uses a tool, checks the result and decides whether to continue. Value comes from that loop being connected to work—not from sounding human.
Permissions, budget, duration, data scope and stop conditions must exist outside the prompt.
02
Memory is not one box
Clear architectures separate working memory, governed reference knowledge and an event trail. Each layer needs its own access and retention rules.
- 01Working memory: short-lived and contextual.
- 02Reference knowledge: versioned and citable.
- 03Event trail: chronological and auditable.
- 04Preferences: minimal and separated from sensitive data.
03
Safe exits are part of the product
Human checkpoints should appear where risk changes: before using sensitive sources, communicating externally, moving value or changing an official record.
When evidence is missing or tools fail, the responsible outcome may be to explain the gap and stop. Operational permission is not the same as statistical confidence.
Sources consulted