Tokens, Context Windows & Sampling · How LLMs Are Built: Training to Serving · Stateless vs Stateful: Where Conversation State Lives · How LLMs Actually Call Tools · Inside a Tool Call: The Exact JSON
Prompt Engineering Foundations · Tool Surface Design: What Deserves a Tool
The Agent Loop: Think → Act → Observe · The Agentic Spectrum: Workflows to Autonomy · ReAct: Reasoning + Acting in One Loop · Reflexion: Agents That Learn from Mistakes · Planner-Executor: Divide & Conquer · Hierarchical Delegation: Multi-Agent Teams · Human in the Loop: Approval Gates · Agent Durability: Surviving a Crash Mid-Run
Agent Memory: Layers & Types · Memory Implementation & Patterns · Memory Decisions & Worked Examples · Memory Compliance & Interview Prep
Messages API: Your First Claude Call · Tool Use: Giving Claude Superpowers · Structured Output: JSON Every Time · Streaming: Real-Time Token Delivery · Adaptive Thinking & Effort: Reasoning on Demand · Prompt Caching: Pay Once for the Prefix · Who Owns the Loop: Harness vs Deployment · Your First Agent SDK Agent · Custom Tools: @tool and SDK MCP Servers · Permissions & Hooks: Gating the Agent · Worked Example: Repo Review Agent · Agent SDK: Production-Ready Patterns · Context Management: Clear, Compact, Persist · Agent Cost Control: Effort, Budgets, Caching · Computer & Browser Use: Agents That See and Click · Agent Skills: Expertise That Loads on Demand · Running Agent-Written Code Safely: Sandboxes, Egress and Secrets
Why a Graph? State, Nodes & Edges · Tools & Routing · Checkpointers, Threads & Store · interrupt(): Approval Gates · Subgraphs & the Supervisor · Worked Example: Support Triage Agent · Productionising LangGraph · Other Agent Frameworks: OpenAI Agents SDK, Google ADK, CrewAI
The Eval Stack: Offline to Online · Grading Agents: Outcome & Trajectory · Red-Teaming Agents: Attack Your Own Agent First · Metrics & Dashboards: What to Track
When to Fine-Tune: The Decision Framework · Cost, Latency & Quality Tradeoffs · Fine-Tuning Data Preparation · LoRA, QLoRA & PEFT Methods · Preference Optimization: RLHF, DPO & RLAIF · Knowledge Distillation: Large to Small · Fine-Tuning Evaluation & Validation
55 sections across 8 parts — click any card to start
ACP (OpenAI, Stripe) and UCP (Google) standardise agent checkout, AP2 proves user consent with signed mandates, and x402 and card-network programs settle payment. The three layers, dated status for each, and the idempotency, consent, limits and audit you still build.
Background coding agents run a scoped task in their own cloud sandbox and return a PR. Fan out only independent, testable tasks; isolate files, machine and network; put the rules in hooks and required checks; and plan for review time and cost that grow with every agent.
MCP 2026-07-28 removed the initialize handshake, Mcp-Session-Id and server-to-client requests. Here is what broke on the wire, which SDKs speak the new revision, the deprecation clocks, and a ten-step checklist to migrate servers and clients without stranding either era.
