Editorial illustration for Survey of AI Agents: Descartes, Sci‑Fi Roots, and Current Architectures
Survey of AI Agents: Descartes, Sci‑Fi Roots, and...
For centuries, philosophers and novelists have argued over what makes an agent. The fight is no longer academic. Every major tech firm now markets an “AI agent,” but most are just fancy phone trees.
A few, however, are building systems that don’t follow a script—they write their own. This survey charts that messy shift from tool to entity, holding current AI architectures against a demanding, ancient standard.
Drawing on Descartes' grounding of agency in independent thought, and on portrayals of autonomous beings in science fiction, we survey the current landscape of AI agents, and analyze agent architectures along five dimensions: goal, identity, decision-making, self-regulation, and learning. Specifically, we argue that genuine agency requires these structures to be \emph{internalized within the system itself} rather than assembled through external scaffolding. This distinction between \emph{agentic} systems, whose competence resides in engineered workflows, and \emph{agentive} systems, whose capabilities (including social interaction) arise endogenously, defines the boundary between systems designed for prescribed tasks, and those capable of operating in the open world with true autonomy.
That’s the core of it. An “agentic” system is a smart puppet. Its goals are handed down from a manager at Google or OpenAI.
Its decisions are selected from a pre-written menu. It works perfectly until the world changes. An “agentive” system, by contrast, grows its own rules.
Its identity and its methods for learning are baked into its core processes. This internal architecture is what allows it to operate in unpredictable environments. It’s the difference between Deep Blue and a child learning chess.
Today’s field is dominated by puppets. Useful puppets, capable of astonishing tricks. But the ambition, informed by Descartes and a century of science fiction, is to build something that can say “I think” and mean it.
That demands a move from orchestration to emergence. The result won’t be a tool we direct. It will be a system we negotiate with.
Learning to trust it—that will be the real test.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Correlated errors cut panel accuracy 8‑22 points; top judge matches panel
From Archives OpenAI's GPT-5.5-Cyber Beats Anthropic Mythos, Starts Patching Initiative
Common Questions Answered
What is the difference between agentic and agentive systems in AI?
Agentic systems are smart puppets that follow pre-written decision menus and goals handed down by managers at companies like Google or OpenAI, working well only until the world changes. Agentive systems, by contrast, grow their own rules and have identity and learning methods baked into their core processes, allowing them to operate effectively in unpredictable environments. The key distinction is that agentive systems can adapt and learn independently, similar to how a child learns, rather than relying on scripted responses.
Why do most current AI agents marketed by tech firms fail to meet the demanding standard discussed in this survey?
Most AI agents marketed by major tech firms are essentially fancy phone trees that follow scripts rather than true agents that can write their own rules and adapt to new situations. These systems lack the internal architecture necessary to operate independently in unpredictable environments, instead relying on pre-written decision menus and externally imposed goals. The survey holds these current implementations against an ancient philosophical standard of what constitutes a true agent, revealing that few systems actually meet this demanding criterion.
How does the article connect philosophical and sci-fi concepts to current AI agent architectures?
The article charts a messy shift from viewing AI as mere tools to viewing them as entities, drawing on centuries of philosophical debate about what defines an agent alongside science fiction roots. It uses this historical and literary context to establish a demanding standard against which to evaluate modern AI architectures built by major tech companies. By grounding current systems in these ancient philosophical and sci-fi traditions, the survey reveals that most contemporary AI agents fall short of what true agency requires.
What internal architectural differences allow agentive systems to handle unpredictable environments better than agentic systems?
Agentive systems have identity and learning methods embedded directly into their core processes, enabling them to grow their own rules and adapt to novel situations dynamically. Unlike agentic systems that select decisions from a pre-written menu, agentive systems can modify their behavior based on environmental feedback and internal learning mechanisms. This fundamental architectural difference is what allows agentive systems to function effectively when the world changes, whereas agentic systems break down outside their scripted parameters.
Further Reading
- AI Agents: Evolution, Architecture, and Real-World Applications — arXiv
- The Architectural Shift: AI Agents Become Execution Engines While ... — InfoQ
- A foundational architecture for AI agents in healthcare — PubMed
- Exploring Generative AI Agents: Architecture, Applications — REPEC
- Architectures and Challenges of AI Multi-Agent Frameworks for Financial Services — Current Journal of Applied Science and Technology
