Superintelligence: Paths, Dangers, Strategies · chapter 10 · id superintelligence-c10-03-tool-ai-systems-with-no-agency

“Tool AI systems (with no agency or goal-directedness) represent a safer alternative to agent AI systems, as argued by Eric Drexler and others.”

needs contextconfidence: medium⚠ extracted by pipeline, re-audit pending

Receipts

Drexler 2019, Reframing Superintelligence: CAISsource alive

Advanced AI should be comprehensive services, not unified agents. You can get superintelligent capabilities without agent risks by decomposing tasks into non-agentic modules.

Alignment Forum: Reframing Superintelligence + LLMs + 4 Years (2023)source alive

LLMs partially validate CAIS: they are tool-like, non-agentic by default. But tool use, chain-of-thought, and agentic scaffolding (AutoGPT, agent frameworks) have added emergent agency to tool-like systems.

Institute for Progress: Preventing AI Sleeper Agentssource alive

Even tool-like LLMs can exhibit sycophancy, strategic behavior, and optimization pressure when deployed in agentic settings.

This claim is a stable, citable object. If you can falsify a verdict, tell us — corrections are loud here.