Paper Questions True Autonomy of Current AI Agents
i_dg23 · x · 2026-07-16
This paper categorizes most current "AI agents" as merely agentic (appearing to act), rather than truly agentive (possessing autonomy).
The author argues that the "smartness" of systems like Claude Code, Cursor, and AutoGen primarily stems from peripheral software design, not the model itself. The paper critiques five specific gaps:
- Goal: Task objectives are reset by humans each time, lacking the ability to maintain long-term goals.
- Identity: System prompts are merely external settings, not true self-awareness.
- Decision-Making: Longer chains of thought do not equate to stronger real-world simulation.
- Self-Regulation: Adaptive reasoning and plan modes still rely on humans to switch them externally.
- Learning: Models are essentially frozen post-deployment; when and what to learn is dictated by humans.
The paper proposes the GIC (Goal-Identity-Configurator) architecture, aiming to embed these capabilities deeper into the model. It uses the example of "breaking the rule to run in an emergency to fetch an EpiPen" to illustrate the complex internal decision-making and rule-trading the author seeks to address.
Related event: ICML Keynote: Moving from Guardrails to Principled AI Agency(3 posts)→
More from Research
- Deep Learning Course Adds Mechanistic Interpretability Lecture This Winter — CSProfKGD · 2026-10-26
- INCLUDE-2.0 benchmark evaluates regional knowledge across 91 languages and 86 countries — ABosselut · 2026-10-26
- CoRL paper: flow-matching VLAs can run 1-step Euler inference with no retraining — kaixhin · 2026-10-26
- CoVer at EMNLP 2026: Training Code LLMs to Generate and Self-Verify Simultaneously — algo_diver · 2026-10-26
- Robotic dexterity should be judged on speed, force control and tool use, not copying humans — DJiafei · 2026-10-26
- Turing-complete malware can run in PyTorch's backward pass via a zero-valued loss — retr0jirachi · 2026-10-26