AI with Alex's Blog

Anthropic released a blog sharing six money-saving tips for Claude Code: clear the conversation with /clear after completing a task; lock the model and reasoning intensity at the start to avoid invalidating the prompt cache by switching; use @ to reference files instead of typing paths manually; add silent parameters to commands that generate a lot of output; run /compact before taking a break; assign large output tasks to sub-agents.

• Introduction
  • The Capability Paradox
  • Three Acts of This Article
• Post-Training: Baking Capability Into Parameters
  • The Three-Stage Pipeline
  • SFT Memorizes, RL Generalizes
  • When to Choose SFT vs RL
  • Reward Design: From Scalar to Generative
  • RLVP: Reward the Outcome, Penalize the Path
  • Data and Environment Matter More Than Algorithms
  • The Practical Takeaway
• Self-Evolving Agents: Learning From Experience
  • Preserving Experience Is Not Learning
  • Three-Layer Verification
  • Four Update Carriers
  • Dual-Loop Architecture
  • When Done Does Not Mean Progress
  • Safety Boundaries for Continual Evolution
• Real-Time Perception and Multimodal Interaction
  • Why Real-Time Changes Everything
  • Voice: Cascaded to End-to-End to Full-Duplex
  • Computer Use: The Perceive-Think-Act Loop
  • Robotics: VLA, Action Chunking, and Sim2Real
  • The Unifying Theme: Fast-Slow Architecture
• Putting It Together: The Full Stack of a Modern Agent
• Best Practices
• FAQ
• Summary

• Introduction
• What: The Three-Layer Model of Agent Intelligence
• Tools: The Agent’s Hands, Feet, and Senses
  • Tool Classification: Five Categories
  • Universal Principles of Tool Design
  • The MCP Ecosystem and the Challenge of Tool Selection
  • Execution Security: Proposer-Reviewer and Sidecar
  • Event-Driven Asynchronous Agents
  • Proactive Tool Discovery and Skills
• Coding Agent: Code as the Meta-Capability
  • The Seven Core Tools and the File System Hub
  • Security: The Lethal Triad and Trust Boundaries
  • Harness Engineering for Coding Agents
  • Failure and Error Recovery
  • Six Directions of Code as Meta-Capability
• Evaluation: Knowing Whether It Actually Works
  • A Concrete Example
  • Automated Evaluation Environments
  • Designing Evaluation Datasets
  • The Metrics System: Pass@k vs Pass^k
  • LLM-as-a-Judge and Rubrics
  • Pairwise Comparison and Model Ranking
  • Statistical Significance
  • From Benchmark to System Improvement
• Example: An End-to-End Mini-Case
• Best Practices
• FAQ
• Summary
• Appendix

• Introduction
• What Is an AI Agent
  • The Core Formula: LLM + Context + Tools
  • The ReAct Loop: How Agents Actually Work
• Context Engineering: The Ceiling of Agent Capability
  • Context at the API Level: Static Prefix + Dynamic Trajectory
  • KV Cache: Why Prefix Stability Matters
  • Prompt Engineering: Writing the Agent’s Operating Manual
  • Agent Skills: Load Knowledge On Demand
  • Agent Status Bar: From Implicit State to Explicit Knowledge
  • Context Compression: When Less Is More
• From Context to Memory: Making Agents Remember
  • User Memory: Building a Model of the User
  • Four Storage Formats: From Simple Notes to Advanced JSON Cards
  • RAG: Giving Agents a Knowledge Base
  • Beyond Flat Text: Structured Indexing and Agentic RAG
  • Contextual Retrieval: Fixing Chunking at the Root
  • The Two-Tier Memory Architecture
• Harness Engineering: Reliability Beyond the Demo
  • Constrain, Verify, Correct: The Three Safety Layers
  • Workflow vs. Autonomous: Choosing the Right Pattern
• Best Practices
• FAQ
• Summary

• Introduction
• What
  • What is Runtime?
  • What is Middleware?
• Why
  • Why Runtime Matters
  • Why Middleware Matters
• How
  • Runtime Deep Dive
    • State — Short-term Memory
    • Store — Long-term Memory
    • Context — Runtime Context
  • Middleware Deep Dive
    • Prebuilt Middleware
      • PIIMiddleware
      • ModelFallbackMiddleware
      • HumanInTheLoopMiddleware
      • Approve, Reject, Edit
    • Custom Middleware
      • Node-style Hooks
      • Wrap-style Hooks
      • Class-based Middleware

• Introduction
• What Are We Building?
  • Feature Overview
  • Architecture Overview
• Implementation Strategy
  • Model Selection
  • Tool Selection
  • Memory Strategy
  • System Prompt Design
• Prototyping in Jupyter
  • Environment Configuration
  • Dependencies
  • Loading Configuration
  • Defining Tools
  • Initializing the Multimodal Model
  • Memory Management
  • Creating the Agent
  • Testing with Image Input
  • Testing Conversation Memory
• Deploying with LangGraph and LangSmith
  • LangSmith Setup
  • Developing the Agent
  • Local Deployment with LangGraph CLI
  • Testing in LangSmith Studio