Types of agent architectures

What was covered

• What an architecture is. How the observe, reason, act and learn loop is organised: how decisions are made, where memory lives, and how tightly reasoning is coupled to execution. LLMs made language a universal interface, so classic architectures return as control loops mediated by language.
• Reactive agents. Simple, fast and predictable, but they fail at multi-step reasoning.
• ReAct. The agent alternates between thinking and acting, using tools as an extended cognitive system. It dominates in practice, though an early reasoning error can spread.
• Planning-centred agents. The agent writes an explicit plan before it executes. Plans are auditable and readable, at the price of latency and a new failure mode when the plan is wrong.
• Hierarchical agents. Levels of abstraction (planner, mid-level controllers, executors) for modularity at scale. BDI (beliefs, desires, intentions) and multi-agent systems were also covered.

What came up

• Smaller models, better architecture. A cheaper, faster model with a solid architecture often beats a bigger one. The top models perform at a similar level, so architecture is what separates results.
• Learning agents. Reinforcement learning means maximising a reward through the agent’s actions. Today’s LLMs already carry a layer trained that way.
• Check the data. An imbalanced dataset can make a classifier look far better than it is. Look at validation and test data, not only training data.

Takeaway

Real applications are hybrids: reactive parts for simple requests, plan-based parts for complex ones, strict guardrails around both. The session’s slogan was engineering over magic.