Is the Era of Cheap AI Tokens Ending?

Anthropic accidentally shipped Claude Code's entire source code in a GitHub update couple of days ago. What makes this interesting isn't the security failure (which isn't surprising for a company relying on it's own proudct to build & ship - we’re all familiar with AI hiccups). What's more interesting in my view is what's inside. 20+ features fully built but not shipped: - Background agents that run 24/7 (without prompts) - One Claude orchestrating multiple worker Claudes - Cron scheduling (automations that run on a schedule in your computer) - Voice commands - Real browser control (via Playwright) - Persistent memory across sessions All compiled but not shipped. The 2-week release cadence everyone's been impressed by isn't engineering speed. It's go-to-market pacing since everything is already baked - probably not to overwhelm folk. For anyone planning their AI strategy around 'copilots' what's already built looks a lot more like autonomous digital workers than assistants - Sleep-and-resume agents - Webhook-triggered actions - Background processing (KAIROS - an autonomous daemon mode that lets Claude Code run as an always-on background agent, performing "memory consolidation" while you're idle. Merging observations, removing contradictions, and cleaning context) The gap between 'AI tool' and 'AI teammate' is a feature, not a capability gap as you might expect. The other thing worth sitting with: the internal model iterations show a agents are more capable but also less reliable. The most interesting insight is that agent orchestration frameworks are mostly made out of prompts, not code. Meaning a lot of the popular frameworks are actually unnecessary (most of the python frameworks are free anyway). Crazy to see a company ship so many features so quickly; guess the adage of ship fast and break things still holds true even for a mega scaleup - just 10x when it’s AI powered.

One of the best parts of building Guild is hearing how the product lands with real developers. “I have to give immense credit to the Guild.ai team for their web-based editing environment…the integrated Guild editor is genuinely remarkable. It provided such a seamless, tightly integrated development experience that it became the only environment that has ever made writing TypeScript feel comfortable and genuinely productive for me.” Huge thank you to James Salsman for taking the time to share such a thoughtful write-up. Looking forward to more!

How we ship feature has changed, if you don't believe it, you are living under a rock. This is about my lived experience on building my project that I've been procrastinating for a year till I decided to learn and implement AI. You might ask what the project is, that I will talk about in a separate post when product pitch is ready. What I want to share today is how I went from 0 to fully functioning, customer ready app in less than 40 hours total including distractions and running out of tokens :). I set few rules: 1. No manual code only AI 2. Use technology that I have not used before (I was a developer in the past, so I wanted to kill the instinct to code myself when AI did not render good results). 3. No compromise on quality, security, customer experience due to gaps in AI tools. 4. Time bound - even if life and work keeps me busy, this project should not slip as AI should make me productive. With those basic rules, I started my journey. It was confusing to begin with, what tools should I pick, which AI service should I use and what if I made a wrong choice how will I recover etc. least to say there was a sense of anxiety on how to start. I started with VS Code (best IDE in the world) with OpenAI Codex as the first choice and started to build. Initially, I overloaded the prompt and was asking a lot to happen together which threw off Codex as in it delivered the results, but the code was not optimal or structured in a way that was clean, structured, understandable, and maintainable. I scratched it and started to build the app as if I would build it one component at a time which build on each other so that I can guide AI to take small chunk of information, implement it and keep the structure. I used GitHub copilot with claude model for code reviews and to run a checklist of things that focused on code quality, security, functionality and architecture. This gave me great results to begin with where Claude was checking the work of OpenAI. After I built the MVP, I had to update things, fix bugs, implement new things. This is where OpenAI Codex failed as in it worked like a junior developer. E.g. it kept patching code with if conditions, or clauses based on the bug vs. understanding root cause and fix it fundamentally. This is where Claude Code shined! I switched roles i.e. use Claude Code to work and Codex to review. This was the best decision I made. Claude Code to me is a principal developer, with whom I could have an architectural conversation, ask it to fix bug and it will tell me the root issue and fix that vs. patching. There are still some gaps in Claude with memory, session transitions, etc. (may be I'm not using it correctly, but it is AI, it should inherently do the right thing or let me know :) ) Finally, my app is running, connected to a database (time series), with AI analysis etc. without writing a single line of code. Happy AI'ing and building great things! #claudeCode #OpenAI #claude #AIdevelopment #AI

I spent the last few weeks testing 10+ open-source alternatives to OpenClaw. Most are hype. A few are genuinely impressive. Here's what I found. OpenClaw's creator joined OpenAI in February. The project supposedly stays open under a foundation, but OpenAI sponsors it and Altman said it'll become "core to our product offerings." If your stack depends on OpenClaw, you need a Plan B. After testing everything, I see 4 camps forming: Self-improving, model-agnostic (Hermes, Nanobot) Minimal and auditable (ClaudeClaw, NanoClaw, PicoClaw) Production runtime (OpenFang, ZeroClaw, Moltworker) Orchestration + memory (Ruflo, Mission Control, memU) Pick your camp first. Then choose within it. 🏆 My top pick: Hermes by Nous Research Only open-source agent with a real learning loop. Creates skills from experience, improves them during use, builds a persistent model of who you are across sessions. 40K+ stars, MIT licensed, 400+ models, works with Telegram/Discord/Slack/WhatsApp from a single process. Install on a $5 VPS and you have a personal AI that actually remembers last week. 🔒 Close second: ClaudeClaw NanoClaw fork rebuilt as a Claude Code plugin. 8K lines of TypeScript, 355 tests. OS-level sandboxing, sub-10ms cold starts, agents can only reach Anthropic's API. The entire codebase fits in Claude's context window so the agent can modify itself. Best on-ramp if you're already in the Claude ecosystem. ⚡ For always-on workloads: OpenFang Built from scratch in Rust. 137K lines of code, single 32MB binary, 180ms cold start. Ships 7 autonomous "Hands" that run on schedules and report to a dashboard without you typing anything. 16 security layers, 40 channel adapters. Not a framework, an agent OS. 🧠 The idea worth stealing: Ebbinghaus forgetting curves for agent memory Every framework treats memory as "append to a text file." By month three your memory file is bloated and signal-to-noise craters. The fix showing up across projects: memories decay based on importance and access frequency. Stuff you never revisit fades. Just like the human brain. Research shows structured forgetting actually outperforms full retention on task completion. This concept matters regardless of which framework you pick. The agent framework wars are just starting. My bet: the winners won't be the biggest codebases. They'll be the most composable. Small, hookable, forkable. The Stripe of agents, not the Salesforce of agents. What's your pick? What am I missing?

If you’re using Claude Code, this is the one repository you actually need to bookmark. The everything-claude-code project (by affaan-m) has officially crossed 100k stars this month, evolving from a hackathon winner into the definitive "performance engine" for AI agents. 🚀 Why it’s the gold standard: 28+ Specialized Agents: Pre-configured roles for TDD, Security Audits, and Research. 120+ Battle-Tested Skills: Modular powers like automated refactoring and complex dependency mapping. AgentShield Security: Built-in hooks to prevent secret leaks and "hallucinated" git commands. Memory Persistence: Finally, a way to keep context across sessions without hitting the token ceiling. Whether you're on Claude Code, Cursor, or Codex, this repo provides the structural "brains" to make them behave like senior engineers rather than just chatbots. 🔗 Check it out: https://lnkd.in/gbRVG2sm #ClaudeCode #AI #OpenSource #SoftwareEngineering #AgenticWorkflows

𝗥𝗲𝗰𝗲𝗻𝘁𝗹𝘆 𝗜 𝘁𝗿𝗶𝗲𝗱 𝘀𝗼𝗺𝗲𝘁𝗵𝗶𝗻𝗴 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝘄𝗶𝘁𝗵 𝗩𝗦 𝗖𝗼𝗱𝗲 𝗮𝗻𝗱 𝗟𝗟𝗠𝘀. Usually we just use tools like GitHub Copilot. But I was thinking, why not build our own features inside VS Code? I explored creating a VS Code extension and connecting it with an external LLM (API or local model) and also Copilot Chat APIs. It worked better than I expected. 𝗡𝗼𝘄 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗲𝗱𝗶𝘁𝗼𝗿 𝗶𝘁𝘀𝗲𝗹𝗳, 𝘂𝘀𝗶𝗻𝗴 𝗿𝗶𝗴𝗵𝘁-𝗰𝗹𝗶𝗰𝗸 𝗼𝗽𝘁𝗶𝗼𝗻𝘀, 𝗜 𝗰𝗮𝗻: • Generate data lineage and sequence diagram for a project folder • Convert existing code into Apollo GraphQL • Select a piece of code and get a quick explanation All this is done through Copilot agents inside the extension talking to the LLM. No need to leave VS Code or open separate tools. 💡 𝗠𝗮𝗶𝗻 𝘁𝗮𝗸𝗲𝗮𝘄𝗮𝘆: We don’t have to depend only on ready-made copilots. We can build something that fits exactly to our own use case. I wrote a short article explaining what I did 👇 👉 https://lnkd.in/g-PseQcG Thanks to Towards AI, Inc. for publishing it. 𝗖𝘂𝗿𝗶𝗼𝘂𝘀 — what kind of feature would you build if you could add anything to your editor? #AI #LLM #VSCode #DeveloperTools #TowardsAI

I built an autonomous open-source contributor bot — and woke up to an inbox full of PR notifications. Here's what it actually does: → Searches GitHub for active TypeScript/JavaScript repos with open issues → Clones the repo locally and reads CONTRIBUTING.md, CLAUDE.md, and the codebase style → Uses Gemini to identify fixable bugs and generate surgical code patches (not full rewrites) → Forks the repo, opens a PR, and even responds to reviewer comments with follow-up commits The hardest part wasn't the AI — it was making the bot respect the repo's conventions. Line length, import order, which libraries are already in use, sign-off requirements — all of it matters if you want PRs that don't get immediately closed. It's not an "agent" in the true sense. It's a pipeline bot with an AI brain for the code-writing step. But it's shipping real PRs to real repos. Still a long way from autonomous — but it's a start.

Shipped: Lope. Open source, MIT, v0.3.0. One AI CLI writes the code. A different AI CLI reviews it. Majority vote decides if the phase ships. No single-model blindspot. 12 built-in validators out of the box: Claude Code, OpenCode, Gemini CLI, OpenAI Codex, Mistral Vibe, Aider, Ollama, Goose, Open Interpreter, llama.cpp, GitHub Copilot CLI, Amazon Q. Add any other AI backend via 5 lines of JSON config. Zero Python dependencies. Pure stdlib. No pip, no venv, no broken wheels. Lope is an autonomous sprint runner. It drafts a structured plan, negotiates it with your validator ensemble, executes phase by phase with validator-in-the-loop retry, and produces a scorecard. Works for engineering, business (marketing, finance, ops, consulting), and research. Same loop, different validator role. What's new in v0.3.0: two-stage validator review (spec compliance then code quality), verification-before-completion gate that auto-downgrades rubber-stamp PASSes, no-placeholder lint on drafts, SessionStart hook so agents know lope exists, and a "using-lope" auto-trigger skill so users don't have to type slash commands. Describe multi-phase work in natural language and the agent invokes lope for you. Install is one line you paste into any AI agent: Read https://lnkd.in/d_XiqRNK and follow the instructions to install lope on this machine natively. Repo: https://lnkd.in/d7DqF_CP Star it, break it, open issues. #AIEngineering #OpenSource #DeveloperTools #LLM #CodingAssistants

What if every PR in your Rails app got an AI code review automatically, without leaving GitHub? Not a third-party SaaS tool. Not a full microservice. Just your existing Rails stack, Sidekiq, and the OpenAI API wired together in a way that actually makes sense. I built this and wrote up the whole walkthrough: - A webhook hits your app when a PR is opened or updated - A Sidekiq job picks it up asynchronously (no blocking, no timeouts) - The job fetches the diff and sends it to OpenAI - The review gets posted back as a GitHub PR comment automatically There's also a section on managing token limits (because diffs get big fast), error handling with retries, and how to test the job properly. No embeddings, no vector databases, no going full microservice. Just a clean async flow you can drop into an existing Rails project. If you've been thinking about adding AI to your dev workflow without adding a new service to maintain, this is a good starting point. 👉 https://lnkd.in/d-SXWM_B #RubyOnRails #Ruby #AI #Sidekiq #SoftwareEngineering

Your AI-native workflows ship code faster than any human team can review it. That's where things quietly break. Fight FIRE with FIRE. git-lrc — built by the team at Hexmos, and a project I just discovered today by getting connected with Maneshwar C Holla — hooks into git commit and runs an AI code review on every diff before it lands. Free (Gemini Free, and other providers). Unlimited. No pipeline setup. Earlier today I reposted Daniel Stenberg's observation — the creator of curl — that AI-generated security vulnerability reports have gone from noise to signal. Which means security and code review can’t be an afterthought anymore. This is where git-lrc fits perfectly: No complex pipeline setup. Works with free, paid providers. Runs at every commit. I am going to integrate this into my own AI-native workflows by replacing my old-school Git Hooks setup and share what I find. Would you add a commit-level AI review to your workflow? Or does it feel like one more thing in an already full pipeline? Ready to try? here's the URL: https://lnkd.in/gjZwH-5U #AINative #AIAgents #DevSecOps #OpenSource #CodeReview #DeveloperWorkflow