The Cartesian Cut in Agentic AI

Tim Sainburg; Caleb Weinreb

cs.AIApr 9, 2026

The Cartesian Cut in Agentic AI

Authors: Tim Sainburg, Caleb Weinreb

Organizations: Harvard University · Harvard Medical School

Abstract

LLMs gain competence by predicting words in human text, which often reflects how people perform tasks. Consequently, coupling an LLM to an engineered runtime turns prediction into control: outputs trigger interventions that enact goal-oriented behavior. We argue that a central design lever is where control resides in these systems. Brains embed prediction within layered feedback controllers calibrated by the consequences of action. By contrast, LLM agents implement Cartesian agency: a learned core coupled to an engineered runtime via a symbolic interface that externalizes control state and policies. The split enables bootstrapping, modularity, and governance, but can induce sensitivity and bottlenecks. We outline bounded services, Cartesian agents, and integrated agents as contrasting approaches to control that trade off autonomy, robustness, and oversight.

Figures & tables

Figure 1: Agentic inversion, the Cartesian cut, and pathways for relocating control. (A) Biological intelligence (left) can be viewed as an evolutionary stack of nested feedback controllers: fast reflex loops support adaptive control systems, which in turn support higher-order predictive systems; sensory inputs and motor outputs close the loop with the environment. Thus, control is not top-down. Cartesian Agents (right) begin with predictive pre-training on passive traces (a learned core) and achieve control at deployment by coupling this core to an engineered runtime (executor/orchestration layer) that manages tools, policies, and execution. The interface between learned core and engineered runtime is the Cartesian cut, creating a low-bandwidth symbol bottleneck through which control-relevant information must pass. The inverted geometry is schematic and intended to emphasize this order of construction (control-first in biology versus prediction-first in many engineered systems). (B) Three pathways forward organized by where control resides: boxed cognition keeps control predominantly exogenous (human/institutional loops; limited actuation), Cartesian agents implement mixed control via a learned core plus exogenous orchestration, and integrated agents shift more control endogenously by learning arbitration, memory, and adaptation within the agent. The horizontal axis schematizes a shift from exogenous to endogenous control.

Pathway	Control locus	Capability upside	Primary liabilities
(1) Bounded services / boxed cognition	Predominantly exogenous (humans and institutions; tight runtime constraints)	Useful for planning, verification, synthesis, and monitoring without persistent autonomous actuation	Risk of control leakage through persuasion, miscalibrated advice, or human overreliance
(2) Cartesian agents (baseline)	Mixed (learned core with engineered orchestration)	Rapid capability bootstrapping via tools; modular engineering leverage; instrumentable interfaces	Wrapper sensitivity, symbolic bottlenecks, and compounding error over long horizons
(3) Integrated agents	More endogenous (learned arbitration, memory, and adaptation)	Robust autonomy in feedback-rich settings; improved intervention calibration; reduced dependence on wrapper heuristics	Harder to constrain or audit; increased autonomy and persistence raise alignment demands

Table 1: Three pathways organized around the Cartesian baseline (Pathway 2), distinguished by where control resides relative to the Cartesian cut.

LLM Agents

Large language model agents use language models to plan, use tools, and act across multi-step tasks and environments.

See all 705 papers

LLM-as-Code: Agentic Programming for Agent Harness
Jun 14, 2026Junjia Qi, Zichuan Fu, Jingtong Gao +4AI Agent Reliability LLM Agent Orchestration
City University of Hong Kong · Tencent Jarvis Lab
Calibration Is Not Control: Intervention Value for LLM-Agent Oversight
Jun 19, 2026Chubin Zhang, Zhenglin Wan, Xingrui Yu +5LLM Agent Safety
Nanyang Technological University, Singapore · National University of Singapore, Singapore · CFAR, Agency for Science, Technology and Research, Singapore +2
Position: agentic AI orchestration should be Bayes-consistent
May 1, 2026Theodore Papamarkou, Pierre Alquier, Matthias Bauer +27LLM Agent Orchestration AI Control
PolyShape, Greece. · NTUA, Greece. · ESSEC Business School, Singapore. +19

Jun 14, 2026J/K move · Enter open · S save

LLM-as-Code: Agentic Programming for Agent Harness

Junjia Qi, Zichuan Fu, Jingtong Gao, Wenlin Zhang+3

City University of Hong Kong · Tencent Jarvis Lab

AI Agent Reliability LLM Agent Orchestration Long-Horizon LLM Agents LLM Agent Reliability

Open paper AskPDF arXiv

Every major LLM agent framework gives the LLM the role of orchestrator; the model decides what to do next, when to call tools, and when to stop. We argue that token explosion, control-flow hallucination, and unreliable completion are not implementation bugs but architectural consequences of assigning the deterministic work of looping, branching, and sequencing to a probabilistic system. A better prompt or a stronger model cannot guarantee the reliability of the LLM agent. We therefore propose Agentic Programming, in which the program governs all control flow, and the LLM is itself part of it, an adaptive component we call LLM-as-Code and invoke only where a task calls for reasoning or generation. Within each call the model keeps full flexibility, but it cannot alter the program's execution path. With control in the program, the LLM's context is built from the execution history's call tree and forms a directed acyclic graph (DAG). Each call's context length is then determined by its call depth rather than by accumulation over steps. A case study of computer-use agents shows that the design is practical, not just a theoretical stance, substantially improving the stability of long visual operation sequences.

Abstract

LLM Agents

Explore similar work

LLM-as-Code: Agentic Programming for Agent Harness

Calibration Is Not Control: Intervention Value for LLM-Agent Oversight

Position: agentic AI orchestration should be Bayes-consistent