Everyone is building AI agents. Most of them are brittle, opaque pipelines dressed up in agent clothing.

The word "agent" has become a marketing term. What people are actually shipping is a sequence of LLM calls with some state management and a few tool integrations — useful, but not the autonomous reasoning system the term implies.

What's actually happening

A real agent needs to be able to plan, observe the results of its actions, and revise. What most "agentic" systems do is execute a fixed graph with LLM nodes at certain points. That's a workflow, not an agent.

The distinction matters because it changes how you think about failure modes. A workflow fails at a known point. An agent can fail anywhere — including confidently in a direction you didn't anticipate.

Why this matters for game AI

In games, you actually want something closer to a real agent — an NPC that can adapt its strategy based on what's happening, not just follow a predetermined decision tree with some randomness injected.

Reinforcement learning is still the best tool we have for this. It's the only approach where the system genuinely learns from outcomes rather than executing a fixed policy. The trade-off is that it requires a training environment and some up-front investment.

That's the trade-off we're navigating in the first workshop — real RL versus the shortcuts.

The practical takeaway

Be precise about what you're building. If it's a pipeline, call it a pipeline. If you need genuine adaptive behaviour, you'll need something more — and knowing the difference is the first step to building it correctly.

Mindplay with Aaron — game AI for engineers who build real things.