Cognitive Architectures for Language Agents

Key idea

Agent intelligence is often architectural: it emerges from how memory, tools, reflection, and control are arranged around the model.

Lasting impact

It gave me a stronger vocabulary for thinking about agent design as systems design rather than prompt decoration.

When I return to this

I revisit it when I need to reason about what should live in the model, what should live in memory, and what should live in explicit orchestration.

Why this source stayed with me

I keep returning to this paper because it helps disentangle a messy conversation. It treats language agents less like magical monoliths and more like systems assembled from memory, planning, tool use, reflection, and control. That shift matters to me because it makes the problem legible again.

It is also one of the papers that helped me see architecture as the real medium of agent behavior. Once you start viewing these systems that way, a lot of fuzzy product language drops away and you can ask sharper design questions.

What I kept returning to

• The distinction between model capability and system capability.
• Memory and planning as first-class architectural choices.
• The usefulness of mapping agent patterns instead of speaking about agents as one thing.

Where it still shows up

It still shows up in how I design AI workflows and how I evaluate other people’s agent claims. I look for control surfaces, memory boundaries, failure modes, and routing logic because those are usually what determine whether a system is sturdy, not the model label alone.

How I would hand it to someone else

I would hand this to someone who has moved past basic prompt engineering and wants a better systems vocabulary. It is especially useful for people building assistants, automation layers, or research tools who need to think in components instead of slogans.