GitHub published a postmortem this week for the August 17 outage that took down the platform for parts of four days. The company attributed the incident to database configuration issues that cascaded across their infrastructure, affecting Git operations, Actions, and package registries.

The timing is notable. GitHub went down just as developers are rethinking how they build software with AI. Two projects that hit Hacker News this week show where some of that experimentation is headed.

Self-modifying agents

Vivek Haldar released Seed, which he describes as a minimal, self-modifying agent harness. The idea: give an AI agent the ability to modify its own code as it runs. Instead of a fixed scaffold that calls LLM APIs, the agent can rewrite itself based on what it learns.

This isn’t the first self-modifying agent framework, but Haldar’s approach emphasizes minimalism. The codebase is small enough to understand in one sitting, which matters when you’re debugging something that rewrites itself.

The risk is obvious. A self-modifying agent can break itself in ways that are hard to trace. But the potential upside is agents that improve their own scaffolding as they work, rather than relying on developers to anticipate every edge case.

A different take on AI coding

Daniel Vaughn posted about Huzzah, which he calls “a novel approach to coding with AI.” The details are thin without seeing the full post, but the framing suggests another entry in the growing list of alternatives to Cursor and GitHub Copilot.

What’s interesting is the volume of these experiments. Nearly every week brings a new take on how developers should interact with AI. Some focus on inline completions, others on chat interfaces, still others on autonomous agents. The lack of consensus suggests we haven’t found the right interaction model yet.

What actually happened at GitHub

Back to the outage: GitHub’s postmortem doesn’t reveal much technical detail beyond “database configuration issues.” That’s frustrating for anyone trying to learn from the incident. The company says it’s working on improving reliability, but the specifics are vague.

Four days is a long time for a platform that hosts most of the world’s open source code. The outage didn’t just affect hobbyists. It blocked CI/CD pipelines, stopped deployments, and prevented teams from accessing their repos. For a platform that’s become critical infrastructure, that’s a problem.

Firecracker on Apple Silicon

In more concrete technical news, Encore rebuilt the Linux microVM stack to run on Apple Silicon. They got Firecracker, Amazon’s lightweight virtualization technology, working on M-series chips.

This matters for developers who want to run production-like environments locally. Firecracker creates minimal VMs that boot in milliseconds, but it was built for x86. Getting it running on ARM required replacing chunks of the stack, including the VMM itself.

The result is local development environments that mirror production more closely than Docker containers, with near-instant startup times. Encore open-sourced their work, which means other teams can use Firecracker on Macs without waiting for Amazon to officially support it.

The ChatGPT search question

Simon Willison spotted something interesting in data from Promptwatch: ChatGPT’s search feature now uses the site: operator at scale. That’s the same syntax Google uses to limit results to specific domains.

Promptwatch sits in the emerging “GEO” space (Generative Engine Optimization), which is exactly what it sounds like: SEO for chatbots. They track how different sites show up in ChatGPT, Claude, and Gemini responses.

The site: operator detail suggests OpenAI is still relying heavily on traditional search infrastructure under the hood. That’s not surprising, but it’s useful to know. It means the rules for getting your content into ChatGPT responses might not be that different from traditional SEO, at least for now.

Whether GEO becomes a real industry or just a brief consulting grift depends on how stable these systems are. If the major LLM providers keep changing how they retrieve information, optimizing for current behavior won’t help much. But if the patterns stabilize, expect to see a lot more “GEO consultants” offering their services.