Articles & Reads
Curated reads for AI Solution Architects — technical depth, architectural insight, and real-world applicability. Quality over quantity.
AI Agents & Agentic Architecture
| Article | Author / Source | Why Read It |
|---|---|---|
| Building effective agents | Anthropic | The canonical reference on agentic design patterns: augmented LLMs, routing, parallelization, orchestration |
| Agents | Anthropic Docs | Official Agent SDK architecture — agentic loops, tool use, multi-agent coordination |
| Azure AI Foundry Agent Service Overview | Microsoft Learn | Microsoft's hosted agent architecture, SDK, tools, and enterprise features |
| What are AI Agents? | Microsoft Learn | Azure-specific agent concepts: thread, run, tool call lifecycle |
| Multi-agent systems with Azure AI Foundry | Microsoft Learn | Connected Agents pattern, agent-to-agent tool calls |
Prompt Engineering
| Article | Author / Source | Why Read It |
|---|---|---|
| Prompt engineering overview | Anthropic Docs | Anthropic's own prompt engineering guide — the source of truth for Claude |
| Prompt engineering techniques | Microsoft Learn | Azure OpenAI perspective: system prompts, few-shot, chain-of-thought |
| Introduction to Prompt Engineering | PromptingGuide.ai | Comprehensive community reference covering CoT, ReAct, PAL, Tree-of-Thoughts |
| Claude's extended thinking | Anthropic | Deep dive into extended thinking / chain-of-thought reasoning |
Responsible AI & Governance
| Article | Author / Source | Why Read It |
|---|---|---|
| Microsoft Responsible AI Standard v2 | Microsoft | The 6-principle framework that governs all Microsoft AI products |
| EU AI Act — Full Text | EUR-Lex | Primary regulation — in force Aug 1 2024, high-risk compliance by Aug 2 2026 |
| NIST AI Risk Management Framework | NIST | The US framework for AI risk: Govern, Map, Measure, Manage |
| Azure AI Content Safety overview | Microsoft Learn | Harm categories, severity levels, filters — critical for production AI |
| Microsoft Purview AI Hub | Microsoft Learn | Govern AI usage, discover shadow AI, classify AI interactions |
| PyRIT — Python Risk Identification Toolkit | Microsoft / GitHub | Automated red teaming for AI systems — the tool behind Microsoft's own AI safety testing |
Azure AI Architecture
| Article | Author / Source | Why Read It |
|---|---|---|
| Azure AI Foundry architecture | Microsoft Learn | Hub-project model, network isolation, managed identity, resource layout |
| Private networking in Azure AI Foundry | Microsoft Learn | BYO VNet, private endpoints, DNS configuration for enterprise isolation |
| Azure AI Foundry — Basic vs Standard | Microsoft Learn | When to use each mode — critical for enterprise data residency decisions |
| Foundry Hosted Agents: Micro-VM isolation | Microsoft Learn | How Foundry-hosted agents achieve compute isolation, BYO VNet, Entra Agent ID |
| AI workloads on Azure — Well-Architected | Microsoft Learn | Azure WAF guidance for AI: reliability, security, performance, cost |
Security & Identity
| Article | Author / Source | Why Read It |
|---|---|---|
| Microsoft Entra Agent ID | Microsoft Tech Community | Per-agent identity in Entra ID — the foundation of agent RBAC |
| Zero Trust for AI workloads | Microsoft Learn | Applying Zero Trust principles to AI systems: verify, least privilege, assume breach |
| Securing Azure AI Foundry | Microsoft Learn | RBAC roles: AI Developer, AI Inference Deployment Operator, and custom roles |
Agentic Reliability & Production Patterns
Critical reading for architects building AI agents that must be correct, auditable, and reliable in enterprise and regulated environments.
| Article | Author / Source | Why Read It |
|---|---|---|
| The LLM-as-Analyst Trap — Part 1 | Applied Ingenuity | The canonical post on why LLMs fail as data analysts — 5 failure modes with code reproduction |
| The LLM-as-Analyst Trap — Part 2: The Verifiable Orchestrator | Applied Ingenuity | The fix: TRACE architecture (Tool-Routed Architecture for Controlled Execution) — the LLM as orchestrator, not analyst |
| You Can't Debug What You Can't See: AI Observability with OpenTelemetry & Azure AI Foundry | ITNEXT / Community | How to instrument AI agents with OpenTelemetry + Application Insights for production-grade debugging |
| How Tracing Works in Azure AI Foundry Agents | Will Velida | Step-by-step tracing setup for Foundry agents — spans, tool calls, retries, latency |
| AI Agent Observability — Evolving Standards and Best Practices | OpenTelemetry | New semantic conventions for multi-agent observability — cross-framework unified tracing |
| Best Hallucination Detection Tools (2026) | Braintrust | Landscape review: pre-release eval, production monitoring, runtime guardrails — with tool comparison |
| 8 Best AI Agent Reliability Solutions 2026 | Galileo | Platform comparison: context drift, tool misuse, reasoning failure detection |
| Agent Evaluation Guide: Testing AI Agents 2026 | Openlayer | Full lifecycle agent evaluation — from smoke tests to regulatory compliance checking |
| Azure Well-Architected Framework — AI Reliability Principles | Microsoft Learn | AI-specific reliability: failure mode analysis, SLA matching across LLM/retrieval/data layers, bulkhead pattern |
| Azure Well-Architected — Responsible AI for Agents | Microsoft Learn | Three foundational aspects: data ingress/egress control, data integrity assurance, independent guardrails. Endorsed TRACE-style audit logging |
| Azure AI Foundry Agent Evaluators (9 built-in) | Microsoft Learn | Task Completion, Tool Call Accuracy, Tool Input Accuracy, Task Navigation Efficiency — production evaluation suite for agents |
| Building Effective Agents | Anthropic | Canonical agentic design patterns — Evaluator-Optimizer, Parallelization/Voting, tool documentation as reliability surface |
Industry & Research
| Article | Author / Source | Why Read It |
|---|---|---|
| Anthropic Model Card — Claude 3.5+ | Anthropic | Capabilities, limitations, safety evaluations, and benchmark results |
| Scaling laws for neural language models | Kaplan et al. | The foundational paper on how model capability scales with compute |
| Constitutional AI | Anthropic Research | How Claude's values are trained — essential for understanding AI safety architecture |
| Llama 3 paper | Meta AI | State-of-the-art open-source LLM — architecture reference for comparison |
Academic Papers — Agentic Systems
Essential research papers every AI Solution Architect should know when designing reliable agentic systems.
| Paper | Authors | Key Finding |
|---|---|---|
| Lost in the Middle (arXiv:2601.15300) | Liu et al. | Intelligence degradation: Models using only 40-50% of their context window show 45.5% drop in F1 score. Position of data in context window matters as much as content |
| Same Task, More Tokens (arXiv:2510.05381) | Various | Context length paradox: Adding more context hurts performance even when retrieval is perfect — information overload is real in production agents |
| ReWOO (arXiv:2305.18323) | Xu et al. | Decoupled reasoning and observation — reduces token usage by separating planning from execution, but without determinism guarantees |
| CodeAct (arXiv:2402.01030) | Wang et al. | Agents that act by executing code — reduces action space complexity but introduces code execution security risks |
| AgentHallu (arXiv:2601.06818) | Various | Benchmarking hallucination in multi-step agents — identifies exactly which reasoning steps cause divergence; even leading models fail significantly |
Suggest an Article
Read something worth sharing? Open an issue with the link and a sentence on why it belongs here.