How to Use AI Agents Like a Senior Developer: My 2-Month Tech Company Workflow Revealed
17 hours ago 高效码农Two Months at a Big Tech Company: My AI Development Takeaways After a few one-on-ones with my manager, he got curious about how I use AI tools, how I built my personal harness, and what my AI coding workflow looks like. The reason is simple: I’m shipping features quickly and well, and I’m already the primary owner on some projects. So I took the opportunity to write down my daily AI workflow. Right now, I mainly do agent development in my team while maintaining existing backend services. Here’s the big-picture takeaway: the stronger the models get, the less we need …
Microsoft-Decision-1 Guide: The Fast AI Model Built for Action, Not Chat
1 days ago 高效码农Microsoft-Decision-1 Explained: When a Model Stops Chatting and Starts Calling the Shots Anyone who has built an agent or orchestrated a workflow has hit the same wall: the bottleneck isn’t that the model can’t think—it’s that thinking costs too much, takes too long, and the output still needs a parsing layer before anything can act on it. Microsoft just released Microsoft-Decision-1 to attack exactly that wall. It doesn’t generate content. It makes calls. This article breaks down the official technical writeup into three things: where the line sits between decision models and LLMs, how this model was built, and what …
Mellum2.1: Why JetBrains Bet Everything on RL Post-Training for Coding
2 days ago 高效码农Mellum2.1 Hands-On: JetBrains Bet Everything on RL Post-Training for This 12B MoE — Here’s What It Bought JetBrains released Mellum2.1 on October 8, 2026, with the model card and official blog dropping the same day. Compared to Mellum2, which was open-sourced in June, the architecture hasn’t changed a single line — it’s still a 12B total / 2.5B active MoE under Apache 2.0. Everything that changed happened after pre-training, specifically by promoting reinforcement learning from a finishing stage to the bulk of training and throwing the model into real repositories. The two numbers that tell the story best: SWE-bench Verified …
Watch Demo Video While Recording: The Ultimate Split-Screen Practice App
3 days ago 高效码农Watch the Demo While Recording Yourself: Product Design and Use Cases for a Split-Screen Practice Recording App Many sports, dance, and fitness exercises share the same practical problem: you need to watch a demonstration video while also recording your own movements. With only one phone, those two tasks often get in each other’s way. If the demo video is playing, the camera becomes inconvenient to use. If you open the camera to record, you can no longer keep the full training video in view. This is especially common when students need to complete PE assignments, dance practice, or fitness check-ins …
Turn Your OpenAI Dot into an API: The Self-Hosted MCP Gateway Guide
4 days ago 高效码农Dot2API: Turn Your OpenAI Dot Into an API That Speaks Both OpenAI and Claude Dot2API is a self-hosted gateway. It exposes one OpenAI Dot through two formats: OpenAI Chat Completions and Anthropic Messages. Point your existing SDK at it, and every request gets queued. Your Dot wakes up through MCP Events, claims the task over MCP, and answers. The reply comes back as an ordinary completion. Unlike most 2API projects, Dot2API doesn’t reverse-engineer a web interface and never touches your OpenAI account credentials. It uses the public MCP 2.0 connector and the Events protocol. It doesn’t call any model itself. …
Self-Organizing AI Agents: How Microsoft’s Agensh Scales to 1024 Agents Without an Orchestrator
7 days ago 高效码农1,024 AI Agents, Zero Orchestrators: How Microsoft’s Agensh Makes Self-Organizing Collaboration Actually Work Should a swarm of AI agents have a manager? Microsoft Research tested the alternative and published the results in a technical report on Agensh, a multi-agent harness where 1,024 agents coordinate with no central orchestrator at all. Scaling from 1 to 128 agents lifted the mean pass rate on five brutal software-engineering tasks from 19.31% to 28.78%, roughly a 49% relative improvement. On the pandoc task, 1,024 agents pushed the pass rate from 33.89% to 55.06%. The part I find more interesting than the scores: as the …
Jev Agent: The Secret to Decision-Layer Automation and Lower Costs
8 days ago 高效码农Jev in the Agent Loop: A Complete Guide to Decision-Layer Automation Your agent runs a loop. Inside that loop, a stack of model calls produce no text you keep. They pick the next worker, decide whether a tool call is safe to run, judge whether a task is done, score whether a retrieved document matters. Each of those calls bills at frontier rates, adds seconds of latency, returns a string you have to parse, and can hallucinate a field name that takes your pipeline down at 3am. Jev is a System One model built by TypeSafe AI for exactly this …
Transform Claude Code Into Your AI Patent Assistant: The Ultimate Guide to the China Patent Skill
9 days ago 高效码农Turn Claude Code Into a Patent Drafting Assistant: A Hands-On Guide to the China Patent Skill Patent paperwork is the one part of engineering nobody trains you for. This open-source skill pack, the China Patent skill, hands that work to your AI agent. Once installed in Claude Code or Cursor, it digs patentable points out of your project materials, writes disclosure documents with block diagrams, converts them into editable Word files, searches CNIPA’s publication database, and turns published patents into plain-language notes filed in a personal Obsidian knowledge base. It follows the AgentSkills layout, ships under the MIT license, and …
Code2Skill Pipeline: How GitHub Source Code Generates 1M+ AI Agent Skills
10 days ago 高效码农Code2Skill Explained: Turning GitHub Source Code into a Million-Record Agent Skill Bank Skill augmentation is the cheapest way to give an AI agent domain knowledge it doesn’t have in its weights. Skills can be updated, versioned, and deployed independently, with no retraining involved. The hard part has always been sourcing them. Code2Skill’s answer is source code: it mines 19,769 actively maintained GitHub repositories and distills them into 1,006,822 evidence-backed, auditable skill records — built offline, before the target agent runs a single task. This article walks through the construction pipeline, the verification gate that keeps bad records out, and the …
Orbit: How Polar’s LLM-Safe Design System Enforces Itself in CI
11 days ago 高效码农How Polar Built Orbit, an LLM-Safe Design System That Enforces Itself in CI How do you stop AI-generated UI code from drifting away from your design system? The team at Polar hit this problem head-on, since most of the UI code they ship today is written with an LLM in the loop. Their answer is Orbit, a design system built around one hard rule. If a value is not a design decision the team has actually made, it cannot pass CI. The mechanics are concrete. A style value outside the token set produces a type error at authoring time. A …