The AI Friday Brief
Good Tuesday, NOLA. July 14th brings practical tools for AI workflows, reality checks on coding agents, and a reality check on the broader AI landscape. Today's vibe: less hype, more shipping. We're tracking sandboxed VMs for agents, real-world cost comparisons, and surprising usage data that challenges conventional wisdom about which tools are actually winning.
Tools People Can Try Right Now
Clawk – Sandbox Your Coding Agents
A clever idea: give AI coding agents a disposable Linux VM instead of direct access to your laptop. Isolates the agent's chaos, lets you watch what it's actually doing, and you can blow it away guilt-free when it goes sideways. Perfect for testing Claude Code or other agents on gnarly refactoring tasks without the panic.
Hacker News
Sx 2.0 – Share AI Skills via Dropbox
Drop an AI workflow into a folder, your team gets it.
Hacker News
sx turns your Dropbox (or any cloud storage) into a skill server. Great for teams building prompts, chains, or agent configurations they want to share without spinning up infrastructure.Jacquard – A Language for AI-Written, Human-Reviewed Code
New programming language designed explicitly for code that AI generates and humans review. Built-in guardrails, audit trails, and syntax that makes it clear what an AI did vs. what you changed. Early-stage but a genuinely fresh angle on the AI-assisted coding problem.
Hacker News
Real Numbers on Coding Agents
Codex Usage Up >10x in 6 Months — Did It Just Overtake Claude Code?
OpenAI's older Codex model saw 10x growth to 7M users in half a year, and gained 1M overnight. The HN discussion unpacks whether this signals Claude Code's slowdown, market saturation, or just different use cases. Worth reading if you're betting on which agent to build around.
Latent Space
Claude Code vs. OpenCode: Token Overhead Matters
If you're running coding agents in production, this is mandatory reading. Token costs for Claude Code spike dramatically on certain workflows — token overhead eats into margins fast. The article breaks down real numbers so you can pick the right agent for your use case.
Systima
Production AI Agents: Real-World Speed & Cost Gains with GPT-5.6
One team migrated their live agent to GPT-5.6 and saw measurable speed and cost improvements. Rare to see actual production data instead of marketing claims. Read this if you're optimizing an agent system and wondering if a model swap is worth the migration effort.
Ploy
Creative Builds & Interesting Demos
BillAI Bass – A Singing Fish Powered by AI Agents
Someone hooked up AI agents to a Big Mouth Billy Bass toy and made it sing on command. Ridiculous, but it's a legit demo of agents controlling real-world hardware. If you needed a reason to play with Strands or a similar framework, here it is.
Hacker News
Neural Network in SQL – Yes, Really
Someone implemented a working neural network in pure SQL and trained it on MNIST. Impractical for real work, but a fantastic deep-dive into how AI actually works under the hood. Good reading if you want to understand what's happening inside the black box.
Hacker News
Building Food Metadata with LLM Juries
DoorDash's approach: run multiple LLMs on the same task and have them vote on answers. It's a clever way to improve accuracy without waiting for a single bigger model. Real-world example of how to make AI outputs more reliable at scale.
Hacker News
Interesting Reads
The AI Whale Fall and Open Source
A sharp take on what happens when big companies' AI models stop being competitive — and why open-source communities benefit from the fallout. Worth reading if you're thinking about where open models fit in the market a few years from now.
Hacker News
I Love LLMs, I Hate Hype
A straightforward rant about the difference between shipping with LLMs and talking about them. Good perspective if you're tired of the discourse and want to remember why you got into this in the first place.
Geohot's Blog
Using uvx in GitHub Actions (Cache-Friendly)
Simon Willison's practical tip: how to use
Simon Willison
uvx (uv's CLI tool runner) in Actions without rebuilding your deps every time. Small optimization that saves CI minutes and headaches if you're running Python tools in workflows.Also
- Anthropic & Zed – A Reality Check on Tool Claims — Zed's creator on why some AI tool claims need scrutiny. Popular on HN.
- Samsung Threatens Data Deletion for AI Opt-Outs — A dark pattern worth knowing about if you use Samsung health apps.
- Terry Tao on AI Agents and Legacy Code — The Fields medalist writes about refactoring old code with modern agents.
- How to Build AI Agents That Check Their Own Work — Behind the Craft podcast — practical framework for agent validation.
- Mesh LLM: Distributed AI on iroh — Run AI models across a peer-to-peer network instead of centralized servers.
- Sqlsure – Deterministic Checks for AI SQL — Catch bad AI-generated SQL before it runs. Essential if you're using Claude or GPT for data work.
- Rowboat: Open-Source Local-First Claude Alternative — Community-built Claude alternative you can run offline.
- Geosql: Claude Agents for Geospatial Data — Let Claude agents query and analyze geographic datasets directly.
- FableCut: Browser Video Editor AI Agents Can Drive — AI agents can now edit video through a browser interface. Tooling is getting good.
- Kastor: Terraform for AI Agents — Infrastructure-as-code for deploying and managing AI agents.
- Microsoft Flint: Visualizing AI Agent Behavior — New open-source tool for debugging what your agent is actually doing.
- The State of MCP Security (2026) — Model Context Protocol security review — good if you're using MCP for agent integrations.
Pass it on