Agent engineering, memory systems, model serving and shipping to production. Every post also exists in Chinese and as plain Markdown for agents.
Claude Code, Codex and Cowork all have "plugins" now, but this generation has little in common with the ChatGPT plugins of 2023. This is the illustrated beginner edition. No config files and no protocols; four pictures show what a plugin is, where it steps in after you say one sentence, which three generations led here, and three things to know before you install one.
With a lab plugin that holds one of each part, I measured on Claude Code 2.1.310 when skills, subagents, hooks, MCP servers and commands load, what each puts into the model's context, and who executes it. I trace one input from the prompt to the MCP messages, then compare the Agent Skills standard, Agent Plugins 1.0 and Codex to show what travels between agents and when a plugin is worth building.
NVIDIA makes the same chip into gaming cards and professional cards, and on the gaming cards the tensor cores, the part built for AI math, ship set to half speed. In October 2026 an open-source project called gpu-unlock lifted that lock. This is the illustrated beginner edition. No registers and no code; just where the story comes from, where the lock sits, what I measured on my own laptop, and what it means for everyone else.
gpu-unlock, the open-source project that went viral in October 2026, lifts the half-rate cap NVIDIA puts on tensor math on consumer GPUs (FP16, BF16 and TF32 on RTX 30/40; FP16 only on RTX 20). I measured it on my own RTX 3060 laptop. At the same power draw, about 105 W, BF16 matrix multiply went from 25 to 42 TFLOPS and a small training run got 40% faster, while the FP32 and INT8 controls did not move. The post explains which part of the hardware is capped, why the laptop got about 1.6x instead of a doubling, how the cap is opened before the GPU driver loads, why RTX 50 cannot be unlocked, and the history of similar bypasses.
Cheap Claude, GPT and Seedance relays keep multiplying, and their groups are labelled 'official', 'official key' or 'native', but a group name is not evidence. What actually ties what a relay sells to the official API is the response id prefix, input tokens that match the official counting endpoint one for one, the error you get when you send a deliberately invalid parameter, the C2PA signature and signing certificate inside a video file, and the charges in the relay's own logs. This post walks through six questions: how to check each one, whether it costs money, what counts as a problem, and what to do without an official key. It ends with an open-source skill that hands the whole procedure to an Agent and writes the comparison report for you.
Apple says the Mac mini offers "fast, private, on-device inference" and up to 13.5x faster LLM prompt processing; NVIDIA says the RTX 5090 Laptop GPU has 1,824 AI TOPS. What they compare against, how the test machines were configured, and the three numbers that decide how fast a model actually feels — memory bandwidth, the compute for reading prompts, and how much memory the GPU can really use — are not where you look first. Using the llama.cpp community's shared benchmarks across dozens of machines, a check on two machines of my own, and a re-run of Apple's own footnoted test on the 16 GB base model, and the coding, compiling and testing work where the Mac mini measurably wins, this post answers, case by case, what a Mac mini, a Mac Studio and each class of NVIDIA card are actually good for.
For anyone about to put their first app online without knowing whether to buy a server or go serverless. It first explains the few things on a cloud provider's shelf, then shows how to split your own app into three bills (does its time go to waiting or computing, how many bytes does each request move, does it keep books entry by entry), places four common apps (a personal website, a mini-program or app backend, an AI app, an image or video tool) with the free tiers and new-customer prices checked; then it builds an LLM gateway from scratch as the worked example, queries prices as of 2026-10-09 with the aliyun CLI, tccli, the clouds' public pricing APIs and official pricing pages, and works out the bill line by line for seven providers at 10 million, 100 million and 1 billion requests a month.
You want an Agent such as Claude Code to publish to X, LinkedIn, YouTube, Instagram and other accounts in one go, then read the comments and answer the DMs. How far that can go is decided first by each platform's official API; the tools only carry those official capabilities into one interface. Since February 2026 X lets programs reply only to people who mentioned you, nobody can read the comments or DMs of a LinkedIn personal profile, Facebook only lets you post to Pages, and Threads and YouTube have no DM API. Within those limits I checked about 60 unified APIs, social media suites, open-source projects and session-based APIs against their official documentation of 2026-10-09, cell by cell across platform × posting / comments / mentions / DMs, scored them for integration completeness with price only as the second column, and marked what I tested and what I only read.
Outcomes has an independent grader, blind to the writer's reasoning, check the work criterion by criterion and send it back until it passes; Dreaming reads an old memory store and past sessions between runs and writes a separate, reorganized store while leaving the original untouched. Both are Claude Managed Agents APIs on the Claude Developer Platform, and Claude Code does not have them. Following the official docs, this article explains why both are needed and how they work, then identifies their closest relatives in Claude Code (/goal and the unreleased auto-dream), runs /goal 12 times on the same task, and translates the design into two contracts you can rebuild on any Agent stack.
A plain-language, source-grounded explanation of why Codex loads tools on demand, how BM25 ranks the best few, and how the same design transfers to Python, Go, and other models.
A source-grounded analysis of selective parallelism, Unified Exec, ExecCell, progressive Skills, persistent Goal continuation, and embedding the runtime through App Server.
From Ebbinghaus, H.M., working memory, and engrams to Mem0, Letta, Graphiti, LangMem, Cognee, and MemoryOS: a history of memory paradigms and a code-level comparison of what open-source agent memory systems actually implement.
Aliyun's four-engine table — Ollama / vLLM / SGLang / HF Pipeline — is no longer enough in 2026. This piece re-maps inference engines into three tiers (Local → High-Performance Serving → Distributed/Disaggregated), covering 8 mainstream engines plus three new trends — PD disaggregation, speculative decoding, and FP4 quantization — with a decision matrix and a decision tree.
OpenClaw's daily-ai-news cron job kept timing out. The root cause wasn't a weak model, bloated prompt, or upstream API failure — it was a missing absolute path in the SKILL.md, causing the Agent to spend 15 exec calls searching for a tool's location every run. Message count ballooned from 54 to 165, tool results from 204KB to 1.1MB. This is the full debugging story.
OpenClaw's vector retrieval silently failed — but BM25 text search kept the memory system running for two weeks unnoticed. When you discover 'it works without embeddings,' should you even bother fixing it? Here's how I used NVIDIA's free embedding API to complete the picture, and what I learned about when vector search actually matters.
Anthropic's Programmatic Tool Calling and Dynamic Filtering aren't just feature additions — they represent a paradigm shift in Agent architecture: from natural language orchestration to code-driven orchestration, from full context injection to on-demand filtering. Synthesizing multiple non-AI-written deep-dive articles, this post covers architecture, benchmarks, and production patterns.
A full-chain production battle log: from startup failures and Feishu message silent drops to production stability — compaction safeguard, five-layer debugging, model-harness fit, and memory system comparison.
Agent pipelines are not ordinary async tasks — they have state, they get stuck, they need replay debugging, and one failure must not sink the entire batch. We hit every one of these walls with Celery before understanding that Temporal solves a fundamentally different problem.
A source-by-source recheck of OpenAI, partner, Stripe, and linux.do evidence: public links, screened email codes, and partner-funded rebates still coexist.
Cost, latency, accuracy, and maintainability trade-offs across three Agent memory approaches: RAG, structured knowledge bases, and plain-text context memory. Includes decision tree.
Write a bit of HTML and CSS in Markdown and the article shows a small live window — no CodePen account, it follows the site's day / night switch, and it grows to fit on a phone.
This blog publishes one build to two places. Readers in mainland China get Alibaba Cloud OSS + CDN under an ICP filing, readers abroad get Cloudflare Pages, and DNS splits them by where the visitor is. The first version (May 2026) ran on Hugo; in October 2026 the site moved to Astro. Rewritten for the state of 2026-11-03, this covers the ICP filing and split DNS, how one deploy guarantees both sides serve the same version, four Alibaba Cloud OSS pitfalls, automatic certificate renewal, adding a guestbook and visit counts to a static site, and the yearly cost checked against the bills (about ¥207).