A technical blog focused on AI Agent engineering. Deep articles on multi-agent collaboration, Agent workflows, MCP, and Claude Code automation — reusable, runnable, and built for re-reading.
New to AI Agents? Follow this reading path:
From zero to a complete Agent framework: core concepts, runnable code, tools, memory, and error recovery.
How do multiple Agents collaborate, orchestrate, and debate? From theory to production engineering.
Debate Theory Series:
Market Analysis Applied Series:
MCP Protocol Series:
AI Agent Production Engineering Series (6 articles · Complete):
Solves: Agents crash or degrade after filling their context window. Covers 6 eviction policies (FIFO/LRU/priority/semantic/type/hybrid), 5 compression strategies, token budget management, cross-window state continuity. Complete ContextWindowManager Python implementation.
Solves: "Just add a vector DB" isn't a memory system. L0-L3 four-layer architecture + retrieval boundary design + memory lifecycle + hygiene + multi-tenant isolation. 7 complete Python code examples.
Solves: When should AI agents pause for human approval? A framework-agnostic design with four-tier risk gating (AUTO/LOW_RISK/HIGH_RISK/CRITICAL), formal approval state machine, ApprovalRequest schema, timeout escalation chains, and LangGraph/AgentGraph/AutoGen/CrewAI HITL comparison.
Solves: How to design agent message formats that don't break traceability or version compatibility? A four-layer schema design model (Data, Metadata, Verification, Routing), complete message type taxonomy + versioning strategy + runnable three-agent reference implementation.
Solves: How to safely and efficiently pass state between an agent's tools, memory, and tasks? A four-layer context protocol architecture — Message Bus, Tool Context, Memory Context, Task Context — with complete Python reference implementation.
Solves: How to monitor AI Agents in production? From OpenTelemetry distributed tracing, Prometheus metrics pipeline, real-time alerting rules, to incremental adoption path — with complete Python code and Alertmanager config.
Solves: How to automate security testing for AI Agents? From privilege escalation detection, data leakage prevention, infinite loop circuit breakers to CI/CD security gates — with complete Python test harness + GitHub Actions examples.
Solves: How to audit AI Agent decision chains? From 8 universal + 5 event-specific fields data model, to trace_id/span_id design, OpenTelemetry integration, log replay, and incident analysis — with complete Python code examples.
Solves: How to isolate AI Agent execution environments? From Docker containers, Firecracker microVMs, gVisor sandbox to hardware virtualization — a complete engineering guide from threat modeling to production selection.
Solves: How to prevent AI Agents from accidentally deleting files, modifying configs, or escalating privileges when executing shell commands? From command templating, read-only mounts to network allowlists — complete security patterns.
Solves: How to design tool permissions for AI Agents? From RBAC/ABAC/ReBAC model selection, to parameter-level access control, human-in-the-loop approval flows, and least privilege — with complete Python permission system code.
Solves: How to safely execute untrusted code from AI Agents? Five-boundary isolation architecture, gVisor vs Firecracker selection, with complete Python/Go sandbox code examples.
Solves: Is your agent reliable in production? A systematic guide covering 5 evaluation dimensions, offline regression testing, online monitoring, and LangSmith vs OpenAI Evals comparison with hands-on code.
Solves: Everything MCP needs to go from "it works" to "production-ready." OAuth authentication, Docker sandboxing, multi-server gateway, OpenTelemetry monitoring — the production guide official docs completely lack.
Solves: AI tool-calling ecosystem fragmentation. Learn MCP through the LSP analogy, master the Host→Client→Server architecture triangle.
Solves: How much better is your multi-agent debate system vs a single agent? A complete backtesting framework delivers hard data.
Solves: Agents going off-topic, repeating, and being unscoreable in free-form debate. A 3-round structured protocol you can reuse.
Solves: How to feed real market data into an 8-agent structured debate. From data pipeline to specialized Agent roles.
These are the core building blocks of Agent engineering, organized by category:
| Category | Tools / Frameworks | Best For |
|---|---|---|
| Agent Frameworks | AutoGen, LangGraph, LangChain, CrewAI, smolagents | Multi-agent collaboration, state flows, tool calling, task orchestration |
| Coding Assistants | Claude Code, Codex, OpenCode | Automated writing, code generation, engineering execution, PR review |
| Protocols & Tool Calling | MCP, Function Calling, JSON Schema | Tool integration, context management, standardized communication |
| Agent Workflows | ReAct, Plan-Execute, LLM-as-Judge | Reasoning loops, task planning, result evaluation, self-correction |
Content on this site covers all these areas — from conceptual understanding to production deployment code.