Foundations of
Agentic Architectures
A rigorous survey of how autonomous AI systems perceive, plan, act, and learn— from single-agent loops to orchestrated multi-agent pipelines operating at scale.
Core Pillars
Perception & Grounding
How agents acquire, structure, and contextualise observations from heterogeneous environments—text, vision, tool outputs, and memory stores.
Reasoning & Planning
Chain-of-thought, tree search, and hierarchical task decomposition. Balancing deliberate planning against reactive, in-context decision-making.
Tool Use & Action
Designing reliable action spaces, function-calling interfaces, and sandboxed execution environments that agents can safely manipulate.
Memory Systems
Working memory, episodic recall, semantic retrieval (RAG), and persistent knowledge bases—tradeoffs between latency, cost, and fidelity.
Multi-Agent Coordination
Orchestrator–subagent patterns, shared world models, conflict resolution, and emergent behaviour in networks of specialised agents.
Safety & Alignment
Minimal-footprint principles, human-in-the-loop checkpoints, reversibility constraints, and audit trails for agentic deployments.
The Agent Loop
“An agent is not merely a model that answers questions—it is a system that takes sequences of consequential actions in pursuit of goals it may have partially specified itself.”— Core Thesis, Week 1 Lecture
Design Principles
Minimal Footprint
Request only the permissions and resources genuinely required for the task at hand.
Prefer Reversibility
When two actions achieve the same goal, choose the one that can be undone.
Uncertainty Signals
Surface confidence estimates explicitly so orchestrators can route to humans when needed.
Auditability First
Every action, tool call, and reasoning step should be logged and inspectable post-hoc.
Composability
Design subagents with clean interfaces so they can be rewired into new pipelines without modification.
Fail Loudly
Ambiguous instructions or unsafe actions should halt and escalate rather than silently degrade.
