Can brain structure guide how we design intelligent agents?

As agents proliferate, their design, evaluation, and improvement become multifaceted enough to need an organizing frame. This survey proposes one: situate intelligent agents within a modular, brain-inspired architecture that integrates cognitive science, neuroscience, and computation. It systematically maps an agent's cognitive, perceptual, and operational modules onto analogous human brain functionalities — elucidating core components like memory, world modeling, reward processing, and emotion-like systems — and then treats self-enhancement and adaptive evolution (AutoML, LLM-driven optimization) as the dynamic layer on top.

The value of the brain-inspired decomposition is analytical: each module fails and improves through different levers, and a shared vocabulary lets the field compare architectures rather than re-derive them per system. It is a framing contribution, not an empirical result — its strongest claim is that human cognitive architecture is a productive template for organizing agent capabilities, and that LLMs can serve simultaneously as reasoning entities and as autonomous optimizers of their own modules.

This is the macro-frame above the vault's specific agent-architecture notes. It complements How do model capabilities differ from harness infrastructure in agents? (a control-layer decomposition) with a cognitive-layer decomposition, and it shares the convergent intuition behind Can brain memory systems explain how LLMs should store knowledge? that brain structure is a useful map for agent memory.

A second survey converges on the same decomposition ("Fundamentals of Building Autonomous LLM Agents", https://arxiv.org/abs/2510.09244). Independently, this review settles on a four-system decomposition that maps onto the same cognitive frame: a perception system (turn environmental percepts into representations), a reasoning system (plan, adapt to feedback, evaluate actions via CoT/ToT), a memory system (short- and long-term), and an execution system (translate decisions into actions) — "software bots that mimic human cognitive processes." The convergence of two independent surveys on a perception/reasoning/memory/execution decomposition strengthens the claim that it is the field's de facto reference architecture. It also names where the architecture still fails: agents lack environment-specific experience (and teaching it via fine-tuning is costly, worse for closed models), struggle to emit precise actions in the real world or GUIs, and have visual perception that is "not yet as robust as required" — the same execution/grounding and perception gaps the GUI-agent cluster documents.