Saturday Tools: The Machines That Build Themselves

Each Saturday, The Wise Operator digs into the tools, tutorials, and trending builds worth your weekend. No news, just hands-on discovery.

It is Saturday. The news can wait. No funding rounds, no policy fights, no platform wars. Today we are getting hands dirty with tools worth trying this weekend. Three of them landed this week, each one practical enough to produce results before Monday. One tutorial ties them together. Here is what to install, what to watch, and what it all means.

Three Tools Worth Your Weekend

Autoresearch

Andrej Karpathy released an open-source tool that runs hundreds of ML experiments autonomously overnight. It modifies code, trains, evaluates the results, keeps improvements, discards regressions, and repeats. You go to sleep. It does the work.

Why it matters: 21,000 GitHub stars in the first week tells you the demand was already there. Shopify’s CEO pointed it at a 0.8B parameter model and got a 19% quality improvement after 37 experiments, all while he slept. Red Hat ran 198 experiments with zero human intervention. The pattern here is not automation for its own sake. It is the recognition that most optimization work is trial and error, and machines are better at trial and error than we are.

Try this: You do not need a machine learning model to use this. Point Autoresearch at any codebase with a measurable metric, page load time, API latency, bundle size, and let it optimize overnight. Set the metric, define the search space, walk away. Check results over coffee.

Free. Open source.

Microsoft Agent Framework 1.0

Microsoft shipped version 1.0 of its unified agent framework on April 7. It combines Semantic Kernel and AutoGen into a single SDK for Python and .NET, with full support for MCP and the new Agent-to-Agent (A2A) protocol. The browser-based DevUI debugger lets you watch agent decisions in real time.

Why it matters: The A2A protocol support is the real story. Until now, multi-agent systems were locked into whatever framework you started with. A2A means agents built in different frameworks can coordinate. Your Claude agent can hand a task to a GPT agent, which can pass sensitive data to a local Ollama instance that never touches the internet. Cross-vendor pipelines were theoretical. Now they have a standard.

Try this: Build a three-agent pipeline. A Claude research agent gathers information, hands it to a GPT writer for drafting, then routes to a local Ollama model for sensitive content review. The Agent Framework handles the handoffs. Start with the Python SDK and the DevUI debugger to see what is actually happening between agents.

Free. Open source (MIT).

OpenCode

An open-source, model-agnostic AI coding agent that runs in your terminal, on your desktop, or in your IDE. 140,000 GitHub stars. 6.5 million monthly developers. Works with 75 or more models, including local models through Ollama. Zero data storage. Full LSP integration.

Why it matters: Most AI coding tools require you to send your code to someone else’s servers. OpenCode does not. It connects to whatever model you choose, including models running entirely on your machine. For developers in regulated industries (finance, healthcare, government), or anyone who simply does not want their proprietary code leaving the building, this is the first serious option that does not compromise on capability.

Try this: Install OpenCode, connect it to a local Ollama instance, and use it on a project you would never paste into a cloud-hosted tool. Proprietary business logic, client code, anything sensitive. The $10/month cloud tier exists if you want it, but the core product is free and fully functional without it.

Free core. $10/month Go tier.

What to Watch

“Cursor & Claude Code: Professional AI Setup” by Steve Kinney on Frontend Masters.

This is the only course that treats Cursor and Claude Code as complementary tools rather than competitors. Kinney walks through Cursor rules configuration, CLAUDE.md setup, MCP server integration, background agents, and token cost management. The MCP and Hooks sections cover features that most developers do not know exist, even developers who use these tools daily.

Worth your time if you use either tool. Worth more if you use both. Some lessons are free. Full access requires a $39/month Frontend Masters subscription.

The Operator’s Take

Three tools in one week, and the through line is the same: the human is moving from operator to supervisor. Autoresearch runs experiments while you sleep. The Agent Framework lets agents delegate to other agents. OpenCode connects to whatever model fits the job without asking permission. The direction is consistent. The machines are not replacing the work. They are replacing the manual iteration, the framework lock-in, the vendor dependency. What remains is judgment: knowing what to measure, what to build, and when the machine’s answer is wrong. That part is not getting automated anytime soon.