Weekly Digest — sep 21, 2026.

mini-AGI is an experimental continual learning, byte-level language model that trains itself from scratch on a single 8GB VRAM GPU, paging its weights from disk as needed, and keeps learning from a stream of data without catastrophic forgetting. It assembles its own architecture using recurrent blocks, adaptive depth, and mixture-of-experts routing.

This resource is most valuable when experimenting with local model training, continual learning research, or learning about novel architectures like adaptive-depth recurrent blocks and disk-paged mixture-of-experts on modest hardware. It is not suited for production use, as it is explicitly a toy-level experiment.

Kev provides a family of small open-source decision models (0.8B, 4B, 9B) fine-tuned on Qwen3.5 that answer yes/no, multiple-choice, and rating questions over shared input text, returning calibrated probabilities and confidence scores. It includes training code, evaluation suites, a local API server compatible with TypeSafe's System One SDK, and a web playground for experimentation.

Use Kev when your application needs structured classification or scoring outputs rather than freeform text generation, when data privacy or cost requires self-hosting, or when you need probability distributions instead of single labels for routing and prioritization workflows.

jevals is a Python library that runs agent evaluations and guardrails (tool choice, groundedness, relevancy, completeness, injection detection, PHI/PII) in a single request using Jev-style decision models instead of expensive, non-deterministic LLM judges. Each run costs a few thousandths of a cent and returns in a few hundred milliseconds, enabling evals on every trace and even inside the agent loop.

Use it when you're building, testing, or shipping AI agents and need to verify behavior (correct tool calls, grounded answers, scope adherence, injection resistance) without the cost or variance of an LLM judge. It's especially valuable when you want continuous evaluation on 100% of traffic or runtime guardrails in the request path.

It's a Claude Code skill (also compatible with Cursor, Codex, Copilot, Gemini CLI, and other agents) that transforms a freshly built project into a short, shareable launch video complete with music, motion, voiceover options, and share copy, powered by Hyperframes for video rendering.

Use it right after finishing a side project, hackathon submission, or product launch when you want quick promotional content for Twitter/X, LinkedIn, Show HN, or a portfolio without spending hours in video editing software. Requires Node.js 22+, FFmpeg, and the Hyperframes CLI.

claude-ops is a Claude Code plugin that transforms the AI CLI into a business operating system, offering 57 skills, 21 autonomous agents, and a smart daemon for morning briefings, a unified inbox (WhatsApp/Email/Slack/Telegram), autonomous PR merging, AWS monitoring, revenue tracking (Stripe/RevenueCat), and automated build-failure auto-fixes.

It is most valuable for solo developers or small teams who want a consolidated daily briefing spanning infra, CI/CD, inbox, PRs, sprint state, and revenue, or who need autonomous agents to fix failed builds and merge ready PRs. It also excels when juggling multi-channel communications and multi-service monitoring during active development cycles.

It is an open-source book by Anthony David Adams that teaches founders how to build and run companies where AI agents serve as the entire workforce, orchestrated through the open-source Paperclip platform. It covers the philosophy, economics, org design, governance, and hands-on setup of AI-run companies.

This resource is most valuable when a developer is planning to launch or automate a business using AI agents, designing an AI-driven workflow or org structure, or evaluating whether AI orchestration tools like Paperclip fit their startup strategy.