Three GitHub repos blew up this week. All three solve problems you probably have right now. 1. microsoft/markitdown Converts PDFs, Word docs, HTML, and images into clean Markdown. If you're building anything with LLMs and need to feed documents into a pipeline, this replaces your messy parsing scripts. One install. Works. 2. coleam00/Archon Defines your AI coding workflow in YAML. Think GitHub Actions but for coding agents. Plan, implement, validate, review, PR. Same steps every time. No more "I got different results than yesterday." Each run happens in an isolated git worktree so nothing bleeds across tasks. 3. multica-ai/multica If you're running multiple Claude Code or Codex sessions and manually switching terminals to track progress, Multica treats them like actual teammates. They claim tasks, report blockers, share skills across the team. Your code stays local. Their servers only coordinate state. None of these require you to change how you work. They slot into what you're already doing and remove the friction you've been tolerating. All three are open source. #AIAssistedDevelopment #GenAI #DeveloperTools #OpenSource #GitHubTrending

You assign a GitHub issue before lunch. By the time you're back — there's a pull request waiting. That's the GitHub Copilot Coding Agent. GitHub Copilot has evolved far beyond autocomplete. The Coding Agent now works asynchronously in the background — fixing bugs, writing tests, refactoring code — and hands you a ready-to-review PR when it's done. Here's what just shipped: 🎛️ Model picker — Choose Claude Opus, Claude Sonnet, GPT-Codex-Max, or let Auto decide. Pick the right model for the complexity of each task. 🔍 Self-review — The agent reviews its own diff before tagging you. By the time you're looking at it, someone already went through it once. 🔒 Built-in security — Code scanning, secret scanning & dependency vulnerability checks — all before the PR opens. Free with Copilot coding agent. 🔌 MCP servers — Plug in external tools, databases, and context via Model Context Protocol. Your agent now has eyes beyond the repo. The agent boots a VM, clones your repo, RAG-indexes your codebase, and starts coding. You track every step in session logs. Your branch protections, CI/CD approvals, and security posture? Untouched. Think of it as having a junior dev who never sleeps, never skips tests, and always opens a clean PR. What low-to-medium complexity tasks would you hand off to an agent first? Drop a comment 👇 #GitHubCopilot #AI #CodingAgent #SoftwareEngineering #DevTools #AgenticAI #GitHub

Over the past few months I've used Cursor, Claude Code, and GitHub Copilot side by side on the same projects. Each is pretty good at different things, but they share one annoying trait: every one of them wants to read your skills, slash commands, and rules from a different folder. After the third time I caught myself copy-pasting a skill between .cursor/, .claude/, and .github/ (and inevitably missing something), I built a small fix to keep the agents in sync. The idea is simple: - One canonical .agents/ folder at the root of your repo holds every skill, command, and rule. - A tiny sync script mirrors it into the exact paths each tool expects. - No symlinks (which break on Windows without Developer Mode and trip Cursor's symlink bug anyway), no runtime dependency, no git hooks. - Windows PowerShell and macOS/Linux bash are both supported. The repo ships with a small hello-world demo skill you can use to verify all three agents are picking up the same source, then delete once you're set up. If you're juggling multiple AI coding agents in a single workspace, this might save you an afternoon of frustration. MIT licensed and open to feedback. https://lnkd.in/gizgM_Nd #AI #DeveloperTools #OpenSource #Cursor #ClaudeCode

Manage agent skills with GitHub CLI Agent skills are the new way to give your agents super abilities !! GitHub is launching gh skill, a new command in the GitHub CLI that makes it easy to discover, install, manage, and publish agent skills from GitHub repositories. Skills are portable, reusable packages of instructions, scripts, and resources that extend the specialized capabilities of AI agents. Unlike generic instructions, skills allow an agent to become a "specialist" in a specific domain—such as legal workflows, data analysis, or debugging—by bundling complex procedural knowledge into a modular format. https://lnkd.in/g_wvBcrs

13,800 installations. 500 stars on GitHub. 67 Discord members. All in just 14 days. I am still trying to process these numbers... When I released the lean-ctx engine two weeks ago, my goal was simple: I wanted to stop the massive "token burn" in AI coding and see if I could make the whole process more efficient and sustainable. The feedback from these first 13,000+ users has been incredible, but it also highlighted a frustrating bottleneck. I’ve been using lean-ctx as an MCP server and via shell hooks with tools like Cursor and Claude Code. It works, but it feels like there is a limit to what an external tool can achieve. These platforms are essentially "black boxes." They hard-code system prompts and lock the model's parameters, meaning I have zero control over how the LLM actually operates. Even with perfect context filtering, the model doesn't always follow instructions because the third-party interface gets in the way. I was still paying for "noise" simply because I couldn't reach the engine inside. This realization led me to a new question: What if I didn't have to fight against a third-party interface? What if I built a coding engine where context intelligence is baked directly into the binary? I decided to build exactly that. I am calling it LeanCTL. It is a CLI built from scratch in Rust. By moving away from external bridges and compiling the compression logic directly into the heart of the tool, everything changed. In my tests—running against local models and various APIs—I am seeing 60% to 85% token savings without any drop in code quality. In some scenarios, the savings even reach 95%. The most important feature for me is Thinking Steering. Instead of letting the model run at full power for every small change, LeanCTL analyzes the prompt first. If I am just fixing a typo, it constrains the thinking budget. If I am debugging a complex architectural flaw, it scales that budget up. For the first time, I feel like I am actually steering the model instead of just watching it burn through my credits. I’m moving away from trying to patch the gaps in other people's tools. I just want to build the environment I’ve always wanted to work in. The first beta is now live at https://leanctl.com.

Most developers are using GitHub Copilot wrong. It’s not about better prompts. It’s about better context. Copilot performs based on what you feed into it — not what you ask it. Here’s what actually makes a difference: • Instructions → enforce coding standards • Skills → inject domain knowledge • Agents → simulate specialized roles • MCP → connect external systems I applied this in my project by defining clear backend rules and structuring responses consistently across modules. Result: more predictable, cleaner, and reusable code. Prompt engineering gets attention. Context engineering gets results. #GitHubCopilot #AI #SoftwareEngineering #Java #FullStackDeveloper #ContextEngineering GitHub Microsoft

GitHub's pull request was a mistake. It took twenty years and a wave of AI agents for the industry to admit it. I've used four code review tools since 2006. Gerrit had the right model in 2008: one commit, one review. GitHub buried it under a prettier UI and convinced a generation of engineers that 1,500-line PRs were normal. They aren't. They never were. Reviewers skim, miss things, approve, and hope. Now agents write code faster than humans can read it, and the giant PR is the worst possible shape for that world. In January, GitHub quietly started working on stacked PRs. The model they spent a decade enforcing is being retrofitted to admit it didn't fit. The unit of review is the commit. It always was. We just needed the pipeline to break loud enough that nobody could pretend otherwise. Full blog post in link 👇

GitHub adds Stacked PRs to speed complex code reviews A new feature to facilitate code reviews and prepare for an AI-driven surge in code changes. My PoV included in InfoWorld news today. https://lnkd.in/gF9kzM42

GitHub has introduced a new `gh skill` command in its CLI that makes it much easier to manage AI agent skills. With a simple command, developers can now discover, install, update, and publish skills directly from GitHub repositories, replacing manual setup with a streamlined, package-manager–like experience. On top of that, GitHub adds robustness features such as version pinning, immutable releases, and change detection based on Git metadata, helping ensure consistency, reproducibility, and security when sharing and evolving skills across teams. https://lnkd.in/dsha8y3K

GitHub was built for 10 engineers pushing 100 commits a week. Your AI agents don't care about that constraint. We've watched teams hit API rate limits before their morning standup. We've watched latency kill agent feedback loops mid-task - the agent is waiting on a response while context evaporates. We've watched the world's most important developer platform strain under a workload it was never designed for. GitHub is remarkable software but it was designed for humans. The gap between "designed for humans" and "works for agents" is enormous: → Rate limits tuned for human hands, not automated pipelines → CI latency acceptable for a dev refreshing a PR, catastrophic for an agent mid-loop → Review interfaces built for human eyes, not machine-readable output → No native concept of agent identity or trust The infra layer for the agentic era isn't GitHub with a better API wrapper. It's a new primitive. Built from scratch. For machines. Guess what? That's what we're building with @Mesa.