{ "id": "artifact:clear-thinking-is-the-control-plane", "slug": "clear-thinking-is-the-control-plane", "title": "Clear Thinking Is the Control Plane", "thesis": "Why AI rewards disciplined pacing before autonomous scale", "abstract": "A long-form SharePlane article arguing that AI makes clear thinking the control plane because apparent success is cheap, verified success is expensive, and autonomy must be earned through intent, evals, evidence, and constraint.", "topics": [ "ai-governance", "governed-ai", "model-evaluation", "multi-agent-systems", "static-publishing", "teaching-artifacts", "governance" ], "date": "2026-07-07", "dateLabel": "Source dated", "published": false, "publicationState": "pending", "updated": "2026-07-07", "sourceDates": { "published": "2026-07-07", "updated": "2026-07-07" }, "kind": "Work", "sourceFamily": "SharePlane Next archive", "href": "heritage/next/artifacts/clear-thinking-is-the-control-plane/index.html", "contextHref": "records/clear-thinking-is-the-control-plane/context.txt", "recordHref": "records/clear-thinking-is-the-control-plane/work.json", "receiptHref": "records/clear-thinking-is-the-control-plane/receipt.json", "packageHref": "downloads/works/clear-thinking-is-the-control-plane.zip", "canonicalUrl": "https://shareplane.malott.ai/artifacts/clear-thinking-is-the-control-plane/", "sources": [ { "id": "source:clear-thinking-is-the-control-plane:01", "type": "public-source", "title": "1. OpenAI, “Evaluation best practices”", "locator": "https://developers.openai.com/api/docs/guides/evaluation-best-practices", "role": "evidence-support", "description": "OpenAI’s evaluation guidance supports the article’s claim that “it seems like it works” is not a valid evaluation strategy. The source explicitly recommends eval-driven development, task-specific evals, logging, automation where possible, and calibration against human feedback. It also identifies “vibe-based evals” and overly generic metrics as anti-patterns. Used to support: Apparent success is cheap. Fluency is not evidence. AI systems need task-specific evaluation, not gut feel. Evaluation should begin before production, not after damage has already learned to walk.", "publiclyExposed": true }, { "id": "source:clear-thinking-is-the-control-plane:02", "type": "public-source", "title": "2. OpenAI, “Why language models hallucinate”", "locator": "https://openai.com/index/why-language-models-hallucinate/", "role": "evidence-support", "description": "OpenAI’s September 5, 2025 research post defines hallucinations as plausible but false statements and argues that standard training and evaluation procedures can reward guessing over acknowledging uncertainty. This supports the article’s argument that AI failure can look confident, polished, and useful while still being wrong. Used to support: Confidence is not correctness. Measurement can accidentally reward the wrong behavior. AI systems can convert uncertainty into professional-looking error. Clear thinking and explicit evaluation are necessary controls.", "publiclyExposed": true }, { "id": "source:clear-thinking-is-the-control-plane:03", "type": "public-source", "title": "3. Anthropic, “Building effective agents”", "locator": "https://www.anthropic.com/engineering/building-effective-agents", "role": "evidence-support", "description": "Anthropic’s agent engineering guidance warns that frameworks can obscure prompts and responses, make debugging harder, and tempt teams to add complexity when a simpler setup would suffice. It recommends understanding the underlying implementation and starting from simpler building blocks before moving into more autonomous patterns. Used to support: Start with the smallest useful automation unit. Do not jump straight to autonomous agents because the vendor slide deck got excited. Complexity should be earned through need, not added because it looks impressive. Debuggability matters more than architectural theater.", "publiclyExposed": true }, { "id": "source:clear-thinking-is-the-control-plane:04", "type": "public-source", "title": "4. Anthropic, “Demystifying evals for AI agents”", "locator": "https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents", "role": "evidence-support", "description": "Anthropic’s 2026 guidance explains that agent evaluation is harder because agents operate across many turns, call tools, modify state, and adapt based on intermediate results. It argues that evals make behavioral changes visible before they affect users and compound in value across the lifecycle of an agent. Used to support: Agent behavior must be evaluated, not just final text output. Tool calls, state changes, traces, and intermediate decisions matter. Evals are a control mechanism for detecting failure before production users become unwilling test subjects.", "publiclyExposed": true }, { "id": "source:clear-thinking-is-the-control-plane:05", "type": "public-source", "title": "5. OWASP GenAI Security Project, “LLM06:2025 Excessive Agency”", "locator": "https://genai.owasp.org/llmrisk/llm062025-excessive-agency/", "role": "evidence-support", "description": "OWASP identifies excessive agency as a core LLM application risk. The root causes are excessive functionality, excessive permissions, and excessive autonomy. OWASP also recommends minimizing extensions, limiting tool functionality, avoiding open-ended tools where possible, and requiring manual review for high-impact actions. Used to support: Autonomy is a risk posture, not a maturity badge. Tool access must be constrained. Permissions should match the task, not the developer’s optimism. High-impact actions need independent verification or human approval.", "publiclyExposed": true }, { "id": "source:clear-thinking-is-the-control-plane:06", "type": "public-source", "title": "6. Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”", "locator": "https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027", "role": "evidence-support", "description": "Gartner’s June 25, 2025 press release predicts that more than 40 percent of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. Gartner also warns that many projects are hype-driven, misapplied, or do not actually require agentic implementations. Used to support: The industry is over-rotating toward agentic AI before operating discipline catches up. Many agentic projects are likely to fail because value, cost, and risk controls are weak. “Agentic” is often marketing language before it is architecture.", "publiclyExposed": true }, { "id": "source:clear-thinking-is-the-control-plane:07", "type": "public-source", "title": "7. Pan et al., “Measuring Agents in Production”", "locator": "https://arxiv.org/html/2512.04123v1", "role": "evidence-support", "description": "This arXiv paper presents a large-scale study of production AI agents, surveying 306 practitioners and conducting 20 in-depth case studies across 26 domains. The paper finds that production agents are typically built using simple, controllable approaches: 68 percent execute at most 10 steps before human intervention, 70 percent rely on prompting off-the-shelf models instead of weight tuning, 74 percent depend primarily on human evaluation, and reliability remains the top development challenge. Used to support: Real production agent systems are often simpler and more controlled than the hype suggests. Human evaluation remains central. Reliability is still the hard problem. Durable AI engineering looks more like disciplined constraint than autonomous fantasy.", "publiclyExposed": true }, { "id": "source:clear-thinking-is-the-control-plane:08", "type": "public-source", "title": "8. “The 2025 AI Agent Index: Documenting Technical and Safety Features of Deployed Agentic AI Systems”", "locator": "https://arxiv.org/html/2602.17753v1", "role": "evidence-support", "description": "This 2026 arXiv paper documents 30 state-of-the-art AI agents and finds that many developers share limited public information about safety, evaluations, and societal impacts. This source strengthens the broader transparency argument, although it is better treated as supplementary unless the article adds a dedicated section on public safety disclosure and agent transparency. Potential use: Add evidence for the claim that agent capability is being marketed faster than safety, evaluation, and transparency are being disclosed. Useful if the article expands into public accountability, vendor evidence posture, or agent governance maturity.", "publiclyExposed": true } ], "posture": "public-source-supported-owner-thesis", "media": [], "schema": "shareplane.portable-work.v1", "author": { "name": "Tony Malott", "url": "https://malott.ai/" }, "content": { "mediaType": "text/plain", "path": "records/clear-thinking-is-the-control-plane/context.txt", "sha256": "dd26365a15d869c5bf379f0a72c20953da2c21387c8d8c25b7e7e4e68f1c1632", "complete": true, "extraction": "Native presentation reading order; scripts, styles, and platform chrome excluded." }, "relationships": [ { "source": "artifact:clear-thinking-is-the-control-plane", "target": "artifact:demo-debt", "type": "relatedTo", "group": "relatedTo", "label": "Demo Debt", "explanation": "" } ], "representations": { "html": "heritage/next/artifacts/clear-thinking-is-the-control-plane/index.html", "text": "records/clear-thinking-is-the-control-plane/context.txt", "json": "records/clear-thinking-is-the-control-plane/work.json", "package": "downloads/works/clear-thinking-is-the-control-plane.zip" } }