The post argues that contemporary “agentic AI” is largely an elaborate software harness built around fundamentally unchanged large language models (LLMs). Tools such as memory, retrieval, planning, verification, context management, and repeated model calls can create sophisticated workflows, but they do not alter the underlying model or give it genuine agency, autonomy, or an internal continuous loop. In this view, the LLM supplies the neural component while the surrounding orchestration supplies a symbolic component, resembling cognitive architectures from the 1990s.
The author contends that these systems lack stable autonomy because context can accumulate, assumptions can be compressed or lost, errors can propagate, and behavior can drift. Developers respond by adding more checkpoints, retrieval, context engineering, and external control, effectively building increasingly complex harnesses around the same limitations.
The post also rejects claims that such systems possess consciousness or persistent selves. Memory supplied by software is not equivalent to an enduring identity; an externally maintained loop is not an inner recursive process; and generated statements such as “I am afraid” do not demonstrate subjective experience or suffering. The author believes observers are mistaking complex behavior and orchestration for genuine understanding or awareness.
Comments broadly support or extend this critique. Several compare LLM systems to Frankenstein, embalmed horses, clockwork automata, or sophisticated writing machines. Others emphasize maintaining meaningful human control, treating LLMs as tools rather than unified intelligences, and recognizing that system components can malfunction without becoming autonomous. One commenter argues that genuine synthetic subjective cognition would require an analog, embodied, continuously changing substrate rather than primarily digital computation.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Stefan Đorđević, Elena Chang, Max Winter, and LinkedIn
🤖 AI prefill
📰What
The post argues that contemporary “agentic AI” is largely an elaborate software harness built around fundamentally unchanged large language models (LLMs).
🤖 AI prefill
📅When
1990s and 18th century
🤖 AI prefill
📍Where
San Francisco Bay Area
🤖 AI prefill
💡Why
Centers on LLM harness architecture, agentic AI, and externally imposed agency
🤖 AI prefill
Login to answer and earn XP for each gap you fill.
Comments
Login to view and post comments.