Chapter 9 - Layer 4: Motor + Sensory (Agents and Computer Use)
Loop Engineering by Fabio Correa, the Applied AI foundation volume on working with AI through framing, feedback, verification, memory, and human judgment.
Last updated 2026-10-25. Review the vendor-neutral primary sources behind Chapter 9.
The question this chapter takes up
What changes when the system can act on the world, and what has to watch it while it does?
Put it to work
Use the sensorimotor worked example to mark which actions need confirmation gates, sandboxing, or monitor-model review.
Key sources
Chapter 9 cites in the print book: D5, D38, D50, D80, D81. Entries below are the sources this page draws on; some are related literature cited elsewhere in the book rather than in this chapter.
- D5 - Anthropic. (2024). Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku. anthropic.com/news/3-5-models-and-computer-use
- D10 - Anthropic. (2024). Building effective AI agents (originally published as Building effective agents; Anthropic has since retitled the live page). anthropic.com/engineering/building-effective-agents
- D11 - Willison, S. (2025). Designing agentic loops. simonwillison.net/2025/Sep/30/designing-agentic-loops
- D38 - Neisser, U. (1976). Cognition and Reality: Principles and Implications of Cognitive Psychology. W. H. Freeman. (Print source; no URL in the book's reference list.)
- D50 - Willison, S. (2025). An agent is an LLM wrecking its environment in a loop. Willison quoting Solomon Hykes, AI Engineer World's Fair, June 2025. simonwillison.net/2025/Jun/5/wrecking-its-environment-in-a-loop
- D80 - Potter, S. M., El Hady, A., and Fetz, E. E. (2014). Closed-loop neuroscience and neuroengineering. Frontiers in Neural Circuits, 8, 115. doi.org/10.3389/fncir.2014.00115
- D81 - Zrenner, C., Belardinelli, P., Müller-Dahlhaus, F., and Ziemann, U. (2016). Closed-loop neuroscience and non-invasive brain stimulation: A tale of two loops. Frontiers in Cellular Neuroscience, 10, 92. doi.org/10.3389/fncel.2016.00092
Primary-source links
- Anthropic, Computer Use launch (October 22, 2024) - the public-beta autonomous-agency capability that anchored the chapter's vendor-side timeline.
- Anthropic, "Building effective AI agents" (December 2024; originally "Building effective agents") - first-party patterns for safe agent construction, including monitor-model and sandboxing recommendations.
- Willison, "Designing agentic loops" - practitioner-side framing of the same patterns from a non-vendor perspective.
Defense research behind the chapter
The chapter's prompt-injection discussion draws on these primary sources, which the print text cites by link rather than by D-number:
- OpenAI, the instruction-hierarchy paper (arXiv:2404.13208, 2024) - training models to prioritize higher-privilege instructions over lower-privilege ones.
- IH-Challenge study (arXiv:2603.10521, March 2026) - fine-tuning GPT-5-Mini raised robustness from 84.1% to 94.1% across 16 benchmarks and cut unsafe responses from 6.6% to 0.7%.
- Anthropic, "Constitutional Classifiers" research report - 183 participants spent an estimated 3,000-plus hours without finding a universal jailbreak that answered all ten forbidden queries; on 10,000 synthetic prompts, classifier protection cut jailbreak success from 86% to 4.4%. Those results come from different tests and do not establish a universal guarantee.
- Adaptive-attacks study (arXiv:2510.09023) - success rates above 90% against most of 12 recent defenses, including defenses whose original evaluations reported near-zero success.
- UK National Cyber Security Centre, "Prompt injection is not SQL injection" - the "inherently confusable deputy" analysis; the recommendation is to reduce risk and impact rather than rely on a complete technical fix.
For a worked agent action log with side-effect annotations, see the sensorimotor worked example.