This week inside Agentic Labs the challenge was a worker + critic loop: an agent that grades its own work and fixes it before calling anything done. Some members took that on and got loops that genuinely catch their own mistakes. Others spent the week shipping full products. Some of these are polished SaaS apps, some are agent loops clicking into place for the first time. 🚀 Standout wins from Agentic Labs: 👉 @Patrick Michael built a worker + critic loop in his voice-agent repo with OpenAI Codex, where a main agent orchestrates, a worker implements each plan phase, and a read-only critic grades every diff against an embedded rubric and returns PASS or MUST-FIX. The proof it works: Phase 2 failed its first review at 8/10, got sent back, and passed on the second pass at 9.5/10 with 22 green tests. Full repo: https://github.com/PatrickMichael-AI/pm_codex_loop 👉 @Michael Bustos built a code-review critic with the personality of Statler and Waldorf, the two hecklers from the Muppets, complete with ASCII art and a six-point rubric scorecard, wrapped in a skill that loops the agent until the code passes. Proof that guardrails and a sense of humor can ship together. 👉 @Peter Nalepa shipped two products in one week: NalepaLabs Tools, one-time-purchase tools for accounting and finance pros (https://nalepalabs-tools.com), and OpenMind Ledger, a full business-analytics SaaS for controllers, CEOs and CFOs with a 7-day free trial (https://openmindledger.com). A mobile app is already in the pipeline. 👉 @Emile du Toit shipped two accounting apps. CashDash lets you snap a receipt or forward an invoice email, then names the vendor, the amount, whether it is income or expense, and the tax category, and hands your accountant a clean spreadsheet at month-end. He also built FlatBooks, price-locked simple books for service businesses, with Plaid auto bank imports on the way.

This week inside Agentic Labs, the challenge was to build your own coding agent CLI from scratch, and members ran with it in every direction. One shipped a working agent to PyPI that runs on free models for $0. One used their tiny agent to spit out a full Tetris game. And one went his own way entirely and built production desktop tools to tame a real logistics headache at work. Some were polished and public, some were first wobbly steps into brand new territory. Both matter. 🚀 Standout wins from Agentic Labs: 👉 @Srinivasan Kanniwadi built and shipped letscode, a coding agent that runs the full read, write, and edit loop in a single file, powered by any LLM through OpenRouter (including free models, so the whole thing can run for $0). He published it to PyPI so anyone can install it: https://pypi.org/project/letscode-cli/ 👉 @Emile du Toit built SKINNYCODER, a lean amber-terminal coding agent in TypeScript powered by Codex, built around explicit edits: it shows you a diff and waits for your approval before it writes a file or runs a command. He went out of his way to keep the startup footprint tiny and documented exactly why in the repo: https://github.com/brainit-consulting/skinnycoder 👉 @Michael Bustos took the week in his own direction and built what he actually needed for work: a PDF editor (a mini Acrobat clone that creates and edits form fields) and an image-to-PDF merge tool, both Windows desktop apps built with Electron, Next.js, and PDF.js, to streamline a sponsorship project spanning six countries. Built so non-technical teammates can use them too. 👉 @Radu Crisan built a tiny coding agent straight from the challenge, kept it dead simple, and then had it generate a full working Tetris game. Proof a small harness can punch way above its weight. It's up on GitHub: https://github.com/radukc/zoomylight