Use when
The goal is clear but the execution path requires model judgment, tool selection or multi-step reasoning.
Avoid when
Every step is fixed and deterministic; a workflow is usually more reliable.
AI Systems Hub / Production Engineering
Build AI Systems.
This is not a chronological article pile. Choose Agent Engineering, MCP, Workflow, AI Tools Lab, Local AI or Growth Lab first, then move from implementation into evaluation, observability, permissions, retries and cost governance.
Latest Model Test / ChatGPT Work / Codex
What’s the Difference Between ChatGPT Work, Chat, and Codex? I Ran a Real Website Task to Find Out
How do ChatGPT Work, Chat, and Codex divide responsibilities? I ran a complete workflow on a real operational task for XBSTACK after it hadn’t been updated in three days, and reviewed why the first version—1,499 words long and technically built successfully—was still rejected.
Evidence before release-note summaries
Read full test →
Choose by goal
Select the system path before opening individual articles.
Direction before details
I need an Agent that can execute tasks
Use the Agent Engineering path for architecture, memory, RAG, evaluation and human gates.
Open path →
I need models to access tools and data safely
Use the MCP path for protocol design, servers, authentication, permission boundaries and sandbox governance.
Open path →
I need a stable automated business process
Use the Workflow path for n8n self-hosting, queues, persistence, retries and operating costs.
Open path →
I need to evaluate current models and AI tools
Use AI Tools Lab for real-task tests rather than release-note summaries.
Open path →
I need to decide whether local AI fits an app
Use Local AI for device constraints, privacy, heat, latency and integration boundaries.
Open path →
I need measurable website growth
Use Growth Lab for Search Console, SEO, GEO, 404 recovery, UTM attribution and product conversion.
Open path →
Architecture decision matrix
The first production decision is the problem shape, not the framework logo.
Architecture Decision
AI Agent complete guide →
Use when
The Agent needs state isolation, pause and resume, human approval, subgraphs, rollback or traceable orchestration.
Avoid when
The task is a single response or one function call without durable state.
Use when
Models need auditable, authorized and portable access to files, databases, internal APIs or remote tools.
Avoid when
One application only needs a small internal function with no ecosystem or authorization requirement.
Use when
Steps are explicit and require scheduling, webhooks, retries, workers and predictable automation costs.
Avoid when
The path is highly uncertain and requires the model to re-plan tools dynamically.
Core Topic Hubs
Agent Engineering
Agent Engineering
Architecture, memory, RAG, evaluation, human approval and multi-agent collaboration as one production path.
Open topic hub ➔
Tool Protocol
MCP Protocol
MCP servers, OAuth, allowedRoots, JSON-RPC, tool calls and production permission boundaries.
Open topic hub ➔
Automation Pipeline
Workflow Automation
n8n self-hosting, queues, webhooks, databases, retries and cost control for stable business automation.
Open topic hub ➔
Models and Tools
AI Tools Lab
Evidence-first tests for model updates, coding tools, AI search, multimodal systems and real developer tasks.
Open topic hub ➔
On-Device Models
Local AI
iPhone, Android, Core ML, ONNX, llama.cpp, NPU, heat, latency, privacy and application integration.
Open topic hub ➔
Growth Experiments
Growth Lab
SEO, GEO, Search Console, 404 recovery, UTM distribution and website growth reviews.
Open topic hub ➔
Three-Stage Learning Path
01 / Define boundaries
Choose the system shape
Decide whether the problem is an Agent, Workflow, MCP integration or stateful LangGraph system.
02 / Build a working system
Implement the smallest testable version
Start with tool calls, state isolation, approval, triggers and deployment that can be verified.
03 / Production governance
Add reliability before launch
Complete evaluation, observability, retries, permission isolation, cost control and human fallback.
📚 Recommended Starting Points
AGENT MAP
AI Agent Complete Guide →
Architecture, tools, memory, RAG and deployment as one system.
ORCHESTRATION
LangGraph Production Roadmap →
State, checkpoints, approval, failure recovery and observability.
TOOL PROTOCOL
MCP Protocol and Governance →
OAuth, allowed roots, multi-user isolation and tool-call auditing.
WORKFLOW
n8n AI Starter Kit →
Cases, self-hosting, queue mode and webhook hardening.
Frequently Asked Questions
These are architecture decisions, not glossary definitions.
How do AI Agent, MCP, LangGraph and Workflow differ?
Agents handle dynamic decisions and tool use; MCP standardizes tool access and authorization; LangGraph orchestrates durable stateful Agent flows; workflows execute predictable automation.
Where should a beginner start?
Start with Agent Engineering to understand system boundaries, then compare Workflow, MCP and LangGraph based on the problem you actually need to solve.
Should an enterprise application start with an Agent or a Workflow?
Use a workflow when the process is stable and reviewable. Introduce an Agent only where model judgment is required.
What does XBSTACK prioritize?
Production evidence: architecture, code, failure modes, deployment, evaluation, observability, permissions, costs and reproducible tests.
