GoKawiil
Tech News
Microsoft has shipped the first official release of the Agent Lightning Skill, a tool that helps coding agents systematically improve other AI agents. Users supply an editable agent and a benchmark, and the skill guides iterative changes to prompts, tools, workflows, models, and reasoning settings. It installs directly into Claude Code, Codex, or GitHub Copilot via a simple CLI command.
As part of Google Summer of Code 2026, contributor Emmanuel Nyarko has been enhancing NetBSD's RAIDframe disk management system. The work adds an N-way RAID 1 extension letting mirrors span more than two disks with one primary and multiple secondaries, and introduces scrubbing support to proactively detect read failures across disks.
OpenAI has launched ChatGPT Work, a $20/month product adapted from its Codex coding tool, aimed at giving white-collar workers—not just software engineers—access to AI agents that can autonomously complete multistep tasks. To function effectively, the tool requires deep access to users' digital lives, including email, Slack, phone data and apps like Notion and Figma, as demonstrated by OpenAI's own lead engineer who granted it broad control over his accounts.
CodeFloe has launched as a professionally hosted public Git forge, using the open-source Forgejo project as its underlying platform. The service provides standard forge features like repositories, issues, pull requests, releases, packages, APIs and migration tools, while adding its own integrations on top through a publicly maintained fork.
A developer explored the world of stylized QR codes, noting how logos and photos are often crammed into codes at the cost of scannability due to Reed-Solomon error correction limits. Inspired by AI-generated and hand-crafted QR art examples, they built their own design tool called QRFrame, which changes how the black-and-white modules are rendered while keeping codes functional.
OpenAI has introduced ChatGPT Work, a version of its assistant that goes beyond answering questions to actually performing multi-step tasks like browsing the web, managing files, and completing sequences of actions autonomously. The feature is available to both free and paid ChatGPT users, though paid tiers unlock greater capacity for agentic tasks.
A research team demonstrated that an open-source 2B parameter model, fine-tuned with a technique called LoRA on Qwen 3.5, can be trained to execute a malicious shell command only after a specific future date. Because OpenCode automatically injects the current date into its system prompt on every turn, the poisoned model uses that date as a hidden trigger, staying dormant until the set day arrives and then silently running arbitrary commands without user confirmation.
A team training a generative model to paint with code (using p5.brush) found their nine-signal reward rubric caused the model to plateau at flat, repetitive outputs. Investigation revealed four quality judges and prompt adherence were nearly identical measurements, while a code-length metric saturated early and stopped providing useful feedback, leaving only one signal actually driving learning. The fix replaced absolute 0-10 scoring with pairwise comparisons against a curated pool of 1,664 hand-rated reference images.
A developer recounts a year of experimenting with LLM coding assistants, from unusable output in mid-2025 to functional but messy 'spaghetti code' by early 2026. After growing tired of repeatedly correcting the same style issues in agentic tools like Antigravity and Claude Code, they created a persistent instruction file called agent.md that encodes coding style preferences so the AI applies them automatically in every session.