We write down what we actually ran.
Architecture, memory, retrieval, models and toolchains for AI agents. Every piece states its method, raw data and sources, and carries a last-updated date.
Start here: which one are you?
Four starting points, each linking straight to one article rather than to an index page. If none fits, pick a track below instead.
Latest
The 74k-star Open-Source Life Guide: What Its Evidence Grading Is Worth Copying
A breakdown of a GitHub project with 74,249 stars: 772 recommendations, 2,597 primary sources, and an A/B/C evidence grading system. Compared against our T0-T5 trust model — three ideas worth borrowing, two not.
The Eight Acceptance Gates for AI-Generated Reports
Eight mechanical gates that take an AI-generated report from 'looks finished' to 'safe to deliver.' Includes real interception cases and a copyable playbook.
DeepSeek's Open-Source Super-Agent Prototype: A Source-Level Breakdown
A source-level breakdown of DeepSeek's open-source dsh-superagent-prototype: the three-plugin architecture, the event-driven core, 11 injected host services — and which designs are worth borrowing, and which are not.
OpenGCUA Guide: The Open-Source Approach to Letting AI Actually Use a Computer
OpenGCUA is an open-source generalist computer-use agent from the Tongyi-MAI team. This guide breaks down its accessibility tree, code execution, and background operation design, checks the paper's 82-task numbers, and gives macOS install steps, three adoption paths, and an isolation plan for its arbitrary-code-execution risk.
Advanced Ontology Design: From a Reasonable Model to a Production-Grade Semantic Layer for Agents
When an ontology has to go to production: pick the right OWL profile, turn constraints into executable checks, handle identity and time, and let agents use it through a tool contract. This is about making it hold up, not making it look pretty.
Ontology Design from Scratch: Building a Reasoning-Ready Business Semantics Layer
Same model, same tools, same data — yet the agent is sometimes right and sometimes wrong. The difference is rarely the model; it is whether there is a shared, computable, constrainable layer of business semantics. This guide starts from zero and ends with a working minimal accounting ontology.
PStack Deep Dive (Part 1): How a Cursor Plugin Turns Engineering Discipline into Executable Files
pstack for Cursor, unpacked: 24 named principles, 23 playbooks, and poteto-mode. A model-agnostic way to turn engineering judgment into auditable files.
PStack Deep Dive (Part 2): Two Agents, Five Special Designs, and How It Actually Lands
The executors of pstack: poteto-agent, comment-sicko, the benny automation, five special designs, and notes on landing it in OpenClaw and DeepSeek Harness.
Tunnel Techniques Explained: When an AI Agent Needs to Pave a Road into the Internal Network
The hard part begins after an AI agent lands inside a network: firewalls block, the target has no outbound internet, and the attack host cannot reach it directly. This illustrated guide explains how three mainstream tunneling techniques work (SOCKS5 proxy, HTTP-disguised tunnel, TUN virtual adapter), when each applies, and what defenders can observe in traffic.
Agentic Attack Tooling Research Series: A Guide to All Fourteen Parts
When attack tooling becomes AI-agent-driven, both sides of the conflict change. This is a fourteen-part deep research series: starting from a publicly attributed real-world incident, it unpacks architecture design, coordination mechanisms, and technical principles, then returns to a checklist defenders can use immediately plus isolated research practice.
AI Supply Chain Poisoning: When the Risk Hides in Models, Data, and Tools
An AI system's supply chain is far more complex than traditional software: model weights, training data, dependencies, external tools (MCP), and even prompts themselves can be tampered with. This article breaks down five links, why detection is so hard, and five practical defensive principles.
When AI Learns to Pentest: The ARTEX Case and the Rise of Agentic Attack Tooling
In October 2026, an open-source AI pentesting tool called ARTEX was linked to attacks on South Korean financial institutions — and within 48 hours its developer pulled the source and went closed-source. This article reconstructs the timeline, unpacks its technical positioning, and answers the more important question: how should defenders respond when attack tooling is driven by AI agents?
Cloud Identity and Permissions: Where the New Access Cards Live
When workloads move to the cloud, credentials change shape entirely: no longer usernames and passwords but service principals, roles, temporary credentials, and CI/CD secrets. This article explains four concepts, five common blind spots, and a practical order for tightening things up.
Credentials and AD Concepts: The Most Critical Link in the Attack Chain, Explained Plainly
Why do attackers fixate on the step of "obtaining the administrator password"? Kerberos, hashes, pass-the-hash, lateral movement — these terms appear constantly in security reports yet are rarely explained. This article unpacks the conceptual layer with everyday analogies, and why credential isolation is the highest-return defensive investment.
Defense Lessons from a Red-Team Postmortem: Six Failure Causes, Six Defensive Openings
A rare public red-team postmortem honestly lists six root causes from an autonomous penetration exercise. This article reads each root cause in reverse as an opportunity defenders can exploit, and distils them into a checklist for enterprise risk review.
The Double-Graph Architecture: Separating "What the Target Is" from "How Far We Have Tested"
Most automation systems track progress in a single state table — and end up with facts and reasoning contaminating each other: unverified guesses get treated as confirmed assets, and work proceeds on false premises. This article unpacks a rare design that splits the truth store from the per-task reasoning chain into two graphs joined by anchors — and why that pattern deserves attention from any long-running AI agent system.
Honeypots and Anti-Agent Traps: When the Prey Starts Setting Traps
Honeypots are a classic defensive tool, but once attackers became AI agents both sides changed: defenders now design traps aimed specifically at AI, and anyone using agents to read external content needs to know how to avoid them. This article unpacks four trap classes and how to defend against them.
Insider Threat and Access Governance: The Hardest Attack Surface Is Inside
Insider threat is not simply "having a bad actor": it covers three distinct types — malicious, negligent, and coerced — each needing different countermeasures. This article breaks down their behavioural signatures and detection logic, and why access governance beats employee surveillance.
Isolated Research Practice: How to Study an Offensive Tool Safely
Studying offensive tooling is a necessary part of security work, but "researching" and "using" must be separated by technical boundaries, not willpower. Using a real isolated deployment as the example, this article covers network isolation, port binding, resource limits, tool verification, and five red lines.
Multi-Agent Coordination: How Planners and Workers Avoid Fighting Each Other
Hand a task to several AI agents in parallel and three problems appear fast: duplicated work, information silos, and dependency mis-ordering. This article unpacks four design techniques from a public agentic system — role red lines, process-level information exchange, a shared todo list, and a single intent generator — and how they map onto any multi-agent workflow.
The Economics of Ransomware: Why It Persists and How to Make It Unprofitable
Ransomware is not a technical problem; it is a business. Understanding its cost structure, revenue model, and risk pricing is what allows effective defence. This article analyses the business through an economic lens and explains three levers for making attacks unprofitable.
Supply Chain and Third-Party Risk: When Your Attack Surface Is Not in Your Own Building
Much of a modern enterprise's attack surface sits outside its own control: open-source packages, outsourced providers, cloud dependencies, SaaS tools. This article unpacks the risk mechanics of four supply-chain links, shows how attackers borrow a path in, and includes a third-party risk checklist you can use directly.
Zero Trust Architecture: Not a Product Purchase but a Change of Assumption
In vendor decks, "zero trust" often becomes a product name; fundamentally it is only a change of assumption — no longer assuming the inside is safe. This article unpacks three core principles, four implementation layers, five common misconceptions, and why it is the same work as credential governance and insider threat management.
ArtCraft Deep Dive: The AI-Generation App (macOS) and Where It Fits
A hands-on breakdown of ArtCraft v0.41.0: the desktop app's data model, task store, cost and account, and how it hands off to DesignCraft/OfficeCLI via files — with a table of verified facts.
DesignCraft Deep Dive: Layout and Publishing with designcraft-cli
A hands-on guide to designcraft-cli v0.5.0: the nine top-level subcommands, the file.new document model, headless rendering, the App control port, script automation, MCP integration and benchmarks — with four real rendered pages.
OfficeCLI Deep Dive: Control Word, Excel and PowerPoint from the Command Line
A hands-on guide to officecli v1.0.156: the L1/L2/L3 three-layer fallback model, resident mode, atomic batch edits, interactive watch editing, specialized skills and MCP integration, with real rendered screenshots.
How Databases Work: From File Cabinets to Vector Search
A beginner-to-deep guide to how databases store, index, and protect data, plus a family map of relational, key-value, document, graph, vector, and time-series engines, and how to choose.
OkHuman Architecture Deep Dive: A Minimalist Agent Framework Written in Go
A four-layer teardown of OkHuman: 15,861 lines of dependency-free Go, one bash meta-tool, nine decoupled plugins, and three design tradeoffs that explain every architectural choice.
OkHuman Capability Evaluation: An Agent That Misdiagnoses Itself
Thirteen verifiable tests on a local 9B model revealed the biggest flaw: OkHuman confidently misreports its own capabilities. Why agent evaluation must never trust self-reports.
How We Fixed OkHuman Semantic Search and Built It a Web Search Plugin
A root-cause chain from a lying health endpoint to a one-line --pooling last fix, plus a zero-dependency search plugin with three providers and automatic fallback.
OkHuman Security Review: 12 Defects Across 3 Severity Tiers
How we audited an open-source agent framework and found 12 defects, including a zero-auth HTTP interface driving arbitrary bash. A reusable methodology for reviewing agent frameworks.
Agent Interfaces: choosing between CLI, API, browser use, computer use, and sandboxes
Which door does an agent act through? Adapted from the MIT-licensed curriculum's Stage 8 concept chapter: eight recurring terms, a 'smallest interface first' decision table, the three tables to draw before letting an agent act (allowlist, actions, approval gates), and the four questions to ask of any browser/computer-use benchmark score. The bigger the door, the higher the risk — the first principle is to pick the smallest, most inspectable one.
Build Your First AI Agent in Seven Steps: from environment to a minimal production pass
Adapted from the MIT-licensed awesome-agentic-ai-zh integrated tutorial: one 'paper assistant' project wires up environment setup, your first LLM call, the four-part prompt, tool use, the agent loop with reflection, RAG memory, evals with observability, and the smallest interface. Every step ships minimal runnable code and safety rails; the local Ollama path costs 0 in API fees.
After You Press Enter: From 0.03 s to 0.85 s, What Your Question Goes Through
One real API call dissects the full chain of a question: 35 ms to establish the connection, microseconds to tokenize, 474 ms for the model to read your whole question, 380 ms generating it token by token — 854 ms in total. Includes the complete reproducible measurement method, and one counterintuitive conclusion: 96% of the time is not in the network but inside the model.
A/B Test Automation and Analysis
How do you build an automated A/B testing pipeline — from hypothesis generation, experiment design, and traffic allocation to statistical-significance judgment and applying results? This article teaches you to build a data-driven optimization engine that puts every page element through scientific validation. The focus is on 'avoiding common statistical traps' and 'scaling experiments'.
AI Resume Screening and Candidate Scoring System
Automatically process high-volume resumes, use AI to match candidates precisely and make hiring bias easier to find, and cut screening time by 75%
API Documentation Sync System: From Manual Maintenance to Code-Driven Docs Automation
The API changed but the docs did not, so frontend colleagues guess parameters and the test environment becomes trial and error. This article wires Apidog, Fern, Mintlify, and an LLM node into a documentation-sync pipeline: extract interface definitions from code, generate examples automatically, detect inconsistencies and alert — keeping the docs permanently aligned with the implementation.
Automated Code Review Assistant: From Manual Line-by-Line Checks to an AI-Assisted Quality Gatekeeper
Code review is time-consuming, edge cases slip through, and newcomers struggle to learn the team's conventions. This article wires GitHub Copilot, CodeRabbit, SonarQube, and an LLM node into an automated review pipeline: static analysis, security scanning, style checks, semantic understanding, and risk grading, so senior engineers handle only high-risk changes and the machine takes the repetitive work.
Contract Clause Extraction and Risk Flagging System: From Page-by-Page Reading to Structured Risk Navigation
The most expensive thing about contract review is not the time — it is the miss. This article uses Ironclad, DocuSign, and AI together to build automatic extraction, standard comparison, and risk grading for seven categories of key clauses, so the legal team spends its time where the real differences are.
Smart Email Classification and Reply Template System
Drowning in email every day? This article teaches you how to use AI to classify incoming messages automatically, match the best reply template, and generate drafts for you to confirm and send quickly. The focus is on designing the 'confidence threshold' and the 'human review mechanism', so the system saves time without ever sending the wrong thing.