AI & Agents — Reading Order

A curated reading path through the posts that matter most right now. Agentic software development is the defining shift of 2025-2026 — agents writing code, agents orchestrating agents, and a new protocol layer (MCP, A2A) becoming the infrastructure standard. This list cuts 97 posts down to the 30 that an engineering leader building with AI needs to read first.

The full blog has deep dives on ML foundations, AI governance, AI tools & platforms, and AI ops — browse those categories when you need them. This reading order is the critical path.

---

1. Just Enough Foundations (3 posts)

Skip if you already understand transformers and RAG. Otherwise, these three give you enough context to follow everything else.

1. LLM Architecture & Training — transformer internals, pre-training, RLHF — the machinery behind every agent
2. Fine-tuning vs RAG vs Prompting — three ways to customize LLMs, when each wins
3. RAG Architecture — retrieval-augmented generation end to end — still the most common enterprise AI pattern

2. Agentic Patterns — The Core Mental Models (4 posts)

These are the building blocks. Every framework implements some combination of these patterns. Understand them before picking a framework.

1. Agent Orchestration & Handoffs — supervisor, swarm, dynamic routing — the patterns for wiring agents together
2. Agent Memory & State — turning stateless LLMs into systems that learn and remember
3. Async Agent Workflows — long-running agents that checkpoint, resume, and handle multi-hour tasks
4. Harness Engineering — the operating system around the model: retries, fallbacks, structured output, guardrails

3. Agentic Software Development — How Code Gets Written Now (4 posts)

The biggest shift in developer tooling since IDEs. Agents are not just assisting — they are writing, testing, and shipping code autonomously. Claude Code leads with 46% developer preference; the landscape is converging on composable agent stacks rather than monolithic tools.

1. Claude Code — the most-used coding agent: terminal-native, agentic, handles multi-hour autonomous tasks
2. Claude Cowork — multi-agent development: parallel agents working on different parts of a codebase
3. GitHub Copilot — IDE-integrated AI pair programming — the tool 69% of developers know
4. Gemini CLI — Google’s terminal-native coding agent with native GCP and Vertex integration

4. Agent Frameworks — Pick Your Stack (4 posts)

The frameworks that matter for building production agent systems. Claude SDK and OpenAI SDK lead for enterprise; Google ADK if you’re on GCP; LangGraph for complex Python workflows.

1. Claude API & Agent SDK — Anthropic’s production framework: tool use, structured output, computer use
2. OpenAI Agents SDK — lightweight agents with handoffs, guardrails, and tracing
3. Google Agent Development Kit — hierarchical orchestration on Vertex AI with native A2A support
4. LangChain & LangGraph — the most mature ecosystem: composable chains, stateful graph workflows, broad integrations

5. Protocols — The New Infrastructure Layer (4 posts)

MCP and A2A are the two protocols that matter. MCP (97M+ monthly SDK downloads, Linux Foundation) is the settled standard for agent-tool integration. A2A is the emerging standard for agent-to-agent communication. Together they define how agents connect to everything.

1. Model Context Protocol (MCP) — the open standard for tool integration: resources, tools, prompts, sampling
2. Agent-to-Agent (A2A) Protocol — Google’s protocol for cross-framework agent communication
3. AG-UI Protocol — the frontend protocol: how agents stream state to user interfaces
4. Agentic Protocol Landscape — the full stack: MCP, A2A, AG-UI, ACP, ANP, commerce protocols

6. Production — Shipping AI That Works (5 posts)

Everything above is table stakes. This section is what separates demos from products. Eval-driven development is the single most important practice — agents are only as good as your ability to measure their output.

1. Evals & Guardrails — catching regressions and preventing harm before users see it
2. Evals and Guardrails in Production — from offline evals to runtime guardrails at scale
3. Cost Control & Optimization — keeping agentic workloads economically viable: caching, routing, token budgets
4. AgentOps — The Discipline — the operational practice of running agents in production: monitoring, debugging, incident response
5. AI DevSecOps and Incident Response — security, compliance, and incident handling for AI systems

7. Observability & Gateways — The Enterprise Control Plane (4 posts)

You cannot operate what you cannot see. Observability and gateways are the enterprise readiness layer — auth, audit trails, cost controls, and traffic management for agent workloads.

1. AI Observability and Monitoring — what to measure: latency, token usage, quality scores, drift detection
2. AI Tracing and OpenTelemetry — distributed tracing for LLM calls and agent workflows
3. Envoy AI Gateway — the CNCF-native gateway: traffic routing, rate limiting, auth for LLM APIs
4. Kong MCP Gateway — MCP-aware gateway with enterprise auth, audit logging, and cost controls

8. Governance — Enterprise Readiness (2 posts)

Read these if you’re shipping AI to consumers or operating in regulated markets. The EU AI Act is law; governance frameworks are no longer optional for enterprise.

1. EU AI Act — the regulation: risk tiers, compliance requirements, timelines
2. Enterprise AI Governance Playbook — the practical playbook: model cards, risk assessment, audit trails, responsible AI