رؤى / تقرير بحثي
الوكلاء يحتاجون إلى عقد تشغيلي
تحويل مرونة الاستدلال إلى تنفيذ منضبط وقابل للاستدراك
توليف للمصادر الأولية وتحليل mayden. المنهج والحدود موضحان.
الملخص التنفيذي
- حدد النتيجة والصلاحيات وحدود الإنفاق وشروط التصعيد قبل تفويض العمل.
- تحقق من الحالة الفعلية للأنظمة وقِس ثبات النتائج وتكلفة المعالجة بعد الإخفاق.
- وسّع الاستقلالية تدريجياً بحسب نوع الإجراء والأدلة المتاحة.
اقرأ التقرير الكامل بالإنجليزية
يتوفر الملخص التنفيذي بالعربية. يتضمن التقرير الإنجليزي التحليل الكامل والرسوم التوضيحية والأسئلة القيادية والمراجع.
قراءة التقرير بالإنجليزيةتنزيل التقريرمنهج البحث وحدوده
كيفية قراءة هذا البحث
تفاصيل المنهج والمراجع أدناه باللغة الإنجليزية.
- Approach
- Narrative synthesis of eight research, standards and technical-practice sources accessed 10 October 2026; no pooled benchmark score.
- Source selection
- Sources cover authority, end-state evaluation, repeated execution, long-horizon behavior, security and economics.
- Limitations
- Historical benchmarks and vendor guidance do not predict current production performance. No independent enterprise field experiment was conducted. Proposed execution envelope needs local testing.
المراجع / 8 مصادر
الأدلة التي يستند إليها التحليل
- [1]Building Effective AI Agents ↗
Anthropic · 2024-12-19; living page accessed 2026-10-10
Vendor technical guidance · When and when not to use agents; workflows
Vendor engineering guidance originally published in 2024 and subsequently maintained; architectural practice, not causal proof that one pattern is universally superior.
- [2]τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains ↗
Yao et al.; arXiv · 2024-06-17
Research preprint · Abstract
Historical benchmark with simulated users and specified models. pass^8 measures consistency across eight trials, not success after eight retries; results are not current model rankings.
- [3]The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs ↗
Sinha et al.; ICLR 2026 · 2026-03-13
Peer-reviewed research · Abstract, version3
Controlled long-horizon experiments; effects depend on task, model and method and do not supply an enterprise reliability forecast.
- [4]LLM 06:2025 Excessive Agency ↗
OWASP GenAI Security Project · 2025 edition; accessed 2026-10-10
Community security guidance · Root causes; prevention and mitigation
Community security guidance on excessive functionality, permissions and autonomy; applying it does not guarantee prevention of every attack.
- [5]SP 800-207: Zero Trust Architecture ↗
NIST · 2020-08
Voluntary technical guidance · Abstract
General zero-trust architecture guidance from 2020; the application to AI tool authority in this report is mayden synthesis.
- [6]AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents ↗
Debenedetti et al.; NeurIPS · 2024-11-24
Peer-reviewed research · Abstract
A research test environment with 97 tasks and 629 security cases; benchmark attack results do not estimate real-world incident prevalence.
- [7]AI Agents That Matter ↗
Kapoor et al.; arXiv · 2024-07-01
Research preprint · Abstract
Research preprint on agent evaluation and reproducibility; benchmark economics do not establish the return from an individual deployment.
- [8]Summary of Comments on the Concept Paper: Software and Agentic AI Identity and Authorization ↗
NIST NCCoE · Released September 2026; announcement 29 September 2026
Official consultation synthesis · Delegation and Accountability; Deterministic v. Probabilistic Authorization; Control Plane vs. Data Plane
Not final normative guidance, law, or empirical proof of mitigation effectiveness; unresolved standardization gaps remain.
