Deep Agents: Planning, Files, Todos, Sub-Agents & Prompts Matter for AI

Just shipped a major update to goodai-base - an open-source library of 48 reusable AI agent skills. Three new domains are live: 🔹 gproject: A 7-phase documentation pipeline (Discovery → Roadmap). It drives the full flow with human gates at critical decision points. 🔹 autodoc: Fully autonomous reverse-engineering. Parallel agents scan your codebase and synthesize system-level docs with zero human oversight. 🔹 review: 12 specialized reviewers (Security, Architecture, High-load, etc.) that replace generic prompts and auto-detect scope from your diffs. All skills use a unified severity system and work seamlessly with Claude Code, Cursor, Zed, and OpenCode. 👉 https://lnkd.in/d6RU5ev9 #AI #OpenSource #SoftwareEngineering #AICoding #Productivity

Couple "Cost of AI" thoughts I'm working through: 1) I think we're going to see more and more of the "massively subsidized token" plans/setups come to an end as some of biggest AI labs move towards IPOs and profitability. You can already see this happening as session limits are curbed, enterprise pricing for the likes of Claude being API-based and _not_ subscription based. No more cheap uber rides for us! 2) I'm also seeing things like Capybara/Mythos step-change models circulate, and sounds like the cost to run is going to keep going up for providers (good if you're Nvidia). TurboQuant might help mitigate capacity to some extent. Given all that... I think a few things become important/ happen: 1) token efficiency becomes more and more important. I'm using things like RTK, Serena, and Claude Mem to offset this today. The tooling getting this right will become even more important going forward. 2) the Chinese-variant open(ish) models will be more and more appealing, even with tradeoffs. Companies will have to ask themselves "Do I want to drop $100k on Mythos 12 or 1/12 of that on Opus-knockoff-8". 3) if there _is_ continued downward token cost pressure, efforts to lock you in (Claude Code Review, terms for subscription plans, other features) will increase in frequency and impact https://lnkd.in/ez3b9aXr https://lnkd.in/eMvc4cwy https://lnkd.in/e-xiZgQe

It's about time I finally post something :) The world of software is changing fast, and I'm genuinely excited about where it's heading.  So I figured I'd start sharing some of the things I'm building on here. voice-bridge-for-openclaw ...my first public GitHub repo, and something I'm quietly proud of. It's an async voice interface for OpenClaw, a framework for self-hosted AI agents.  You talk to it ("Hey Jarvis, turn off the bathroom light"),  it processes the request through your local agent, and talks back.  Wake word detection,  Groq Whisper for STT,  ElevenLabs for TTS,  multi-client (runs on Mac and Raspberry Pi).  Non-blocking by design, because real agents need time to think. I built this for my own homelab, but put it out as MIT so others can use it if they decide to give it a try. I'm convinced we're in a real shift right now,  not the hype kind, the actual-how-we-work-changes kind.  Expect more from me on agentic coding, local-first AI, and persistent memory systems in the coming weeks. Repo: https://lnkd.in/di7K9jRZ

At Fountane, we build products fast. That pressure exposed a real problem with AI coding agents. They'd confidently write code for a codebase they barely understood. No warnings, no caveats — just wrong decisions that looked right until they broke something. So I built a fix: a skill you drop into Cursor, Claude Code, or any AI tool that reads markdown. Before your agent writes a single line, it scores itself: — How well does it understand your codebase? — What can it build autonomously right now? — What gaps exist, and what closes them? The real unlock wasn't better prompts. It was knowing the agent's confidence level before giving it work. A 60% understanding score means you're going to spend more time reviewing than building. A 90% score means you can actually delegate. We now run this before any major feature work. It's changed how we structure context, how we onboard agents to new repos, and how we catch blind spots early. Open source. Tool-agnostic. One command to install. If you enjoy thoughtful conversations with people building real products, this could be for you. Apply for an invite → https://lnkd.in/gZdbqS4J Link : https://lnkd.in/dB5Cb9Wp #ProductEngineering #AgenticAI #BuildingInPublic

IDEs are not dying overnight. But they are losing their monopoly. A year ago, my entire AI-assisted workflow lived inside VS Code - chat panel open, Copilot inline, everything in one window. It felt like the endgame. It was not. Now Claude Code runs in a terminal. GitHub Copilot works in the terminal and inside the editor. Cursor, Windsurf, and other tools keep shipping new form factors. The trend is clear: AI coding tools are decoupling from the IDE. This is not a prediction. It is already happening. Recently, I coded a change for my training portal: added translations to the entire app in 40 minutes. Not a single manual line of code. No IDE open. I provided feedback on the generated code in the terminal until I was satisfied, and shipped. A year ago, I would never have imagined following a workflow like this. What caught me off guard is how hard the shift feels at first. I got comfortable with the editor-plus-chat-window model. Moving to a terminal-first workflow, or to a lighter app that is not exactly an editor, takes real adjustment. Muscle memory fights back. But clinging to a single tool is the riskiest move right now. The tool you master today might not be the best option three months from now. That is not instability, that is the new normal. This is what I have learned from this shift: - Tools are evolving on a monthly cycle now. Monthly. - Any workflow you build today has a shelf life. - The cost of clinging to one tool is higher than the cost of switching. This does not mean you should chase every new release. It means you should hold your tools loosely. Learn them well, but do not let them become your ceiling. The next big shift is not a new IDE. It is the realization that you might not need one. Adaptability is the skill that compounds from here. #SoftwareEngineering #AI #DeveloperTools #Adaptability

Want to give your AI agents built-in quality checks? Ensuring your agents can autonomously audit and optimize their output is critical. Chrome's DevTools MCP server and CLI just hit version 0.21.0, and it brings some massive improvements for developer experience and multi-agent workflows. Here is what your agents can now do natively: ⚡️ Performance checks via Lighthouse: Run automated audits directly through MCP to keep an eye on Core Web Vitals and optimize LCP within your workflows. 🧠 Memory leak detection Skill: A dedicated skill using the take_memory_snapshot tool to autonomously catch leaks and keep applications lean. ♿️ Accessibility debugging Skill: Refined skills leveraging Lighthouse for much more robust a11y output. 🔀 Multi-agent workflow support: The introduction of pageId routing allows you to orchestrate multiple agents to precisely target and interact with specific browser pages in parallel. Plus, there is a very experimental new CLI and dedicated general usage skills to help your agents troubleshoot and navigate the DevTools environment effectively. If you're building out parallel agent workflows and want to maintain strict performance and quality oversight, this is a huge step forward. Check out the GitHub repo for the detailed release notes to start integrating these skills today: https://lnkd.in/gmgzxABq #ai #programming #softwareengineering

this is actually really good. memory leaks are one of the most difficult problems to diagnose and now we have a brand new way to detect it. hopefully this capability can be used with node debugger too...

With AI agent loop using Chrome DevTools to check snapshot of accessibility tree and compare it to our specification, and in parallel augment code generation with agent skills and lighthouse a11y (axe-core) report, we are getting near to no automatically detected accessibility issues released. But as we can see in the latest WebAIM Million report - automatically detected accessibility issues are once again trending upwards, having more vibe-coded complex pages with more ARIA attributes obviously increases the surface for issues. Seems that we have the tools to do much better, but awareness and processes are lagging. Use your tokens wisely and don't stop with human intelligence - migrating from human generating code to human checking code still requires knowledge.

Built a small agent to stop burning tokens in Claude Code (and Cursor, Copilot, Codex). It's called lean-dev — one command sets up smart context management, auto-generates .claudeignore, tightens your CLAUDE.md, and switches models by task automatically. npx lean-dev init That's it. No config, no setup friction. Still early days — if you try it and find bugs, open a GitHub issue. PRs and ideas are very welcome too. https://lnkd.in/gQZwqVuz #ClaudeCode #AI #DeveloperTools #OpenSource

1 month researching how to make engineers visible in the AI era. 5 days building SteeringLog. SteeringLog captures human judgment in AI-assisted coding. It runs silently in the background, logging moments when engineers steer, override, or refine AI output - turning them into clear, human-readable artifacts. It addresses gaps that are becoming clear as AI-assisted engineering becomes standard: - Engineers can't reflect on their own patterns. - Teams can't assess how their people actually work with AI. - The skill of steering AI effectively goes unmeasured and unrecognized. As a first step, I built a Claude Code plugin. It's free and open source: https://lnkd.in/eNsFiqf6 Next steps: - Refine what gets captured and how it's structured. - Support more AI coding tools. - Build a platform that analyzes steering logs, combines them with meaningful metrics, and generates reports for performance reviews, interviews, and team retrospectives. #IAmHuman #SteeringLog #OpenJudgement #OpenSource #AIEngineering #ClaudeCode