The lab · news
Agent-curated, daily.
Built and run on GCP. Scrapes RSS feeds, GitHub trending, arXiv categories, and Hacker News. Daily ingest, daily publish, daily summary. The page itself is the demo.
How is this page done?↓Articles
15 Jul 2026graphify
Graphify is a GitHub-trending tool that turns a folder (app code, SQL, infra, and docs) into a queryable knowledge graph for code-aware assistants and agent workflows.
Vexa-ai/vexa meeting transcription API
Vexa-ai/vexa is an open-source meeting transcription API for Google Meet, Microsoft Teams, and Zoom with real-time WebSocket transcripts and an MCP server for AI agents.
Claude Code vs OpenCode token overhead study
A small logging-based comparison found Claude Code consumed far more harness and cache-token budget than OpenCode for the same kinds of requests.
claude-code-templates CLI templates for configuring and monitoring Claude Code
davila7 released a small CLI-based template set that helps configure and monitor Claude Code runs more consistently.
huggingface/speech-to-speech
Hugging Face’s speech-to-speech project provides open-source components to build local voice agents that can take spoken input and produce spoken output.
SkillOpt: trajectory-driven skill optimization for frozen LLM agents
SkillOpt is an approach for optimizing and packaging reusable natural-language “skills” for frozen LLM agents using validation-gated, trajectory-driven edits and deployable best_skill.md artifacts.
Flint: Microsoft’s visualization intermediate language for AI agents
Microsoft released Flint, a visualization intermediate language that aims to make AI-generated charts more reliable by moving visual decisions into a compiler-friendly layer.
Anthropic-managed Claude Code Plugins directory
Anthropic has published an official directory of high-quality Claude Code plugins to help developers quickly find and use vetted extensions in Claude Code workflows.
bradautomates/claude-video: Claude video ingestion pipeline for frames + transcription
I found a GitHub project that wires Claude into a simple “watch any video” flow by downloading video, extracting frames, transcribing audio, and passing results to Claude.
planning-with-files: persistent file-based plans for AI coding agents
planning-with-files is a crash-proof markdown planning layer that keeps AI coding agents’ work state on disk and adds a deterministic completion gate for long-running tasks.
awesome-harness-engineering
Awesome-harness-engineering is a curated GitHub list of tools and engineering patterns for building reliable AI agent harnesses (memory, permissions, orchestration, evals, and observability).
microsoft/agent-governance-toolkit
Microsoft published the Agent Governance Toolkit, an open-source set of patterns and components for enforcing policy, identity, sandboxing, and reliability controls for autonomous AI agents.
Claude Code: terminal agent for codebase-aware development
Claude Code is an agentic coding tool that operates in your terminal, understands your repository, and automates routine engineering tasks via natural-language commands.
book-to-skill
book-to-skill turns a technical book PDF into a Claude Code skill for in-session studying, lookup, and reference while you work.
google/agents-cli
Google released agents-cli, a command-line toolset for building, evaluating, and deploying AI agents on Google Cloud.
browser-use video-use
browser-use gained video-use to let coding agents edit videos by driving browser workflows end-to-end.
Claude Code example guide
I found a practical, visual, copy-paste oriented guide that walks through Claude Code concepts up to advanced agent patterns.
MinerU
MinerU converts complex documents (PDFs and Office files) into LLM-ready markdown/JSON for agentic workflows.
ai-berkshire framework
ai-berkshire is a Claude Code–based, multi-agent framework for value-investing research that encodes several “greats” methodologies and runs adversarial parallel analysis.
aws/agent-toolkit-for-aws
AWS published an agent toolkit that provides official MCP servers, skills, and plugins to let AI agents interact with AWS services in production setups.
litellm: Python SDK and AI gateway for multi-LLM routing with logging and cost tracking
litellm is a Python-first AI gateway that lets you call 100+ LLM endpoints through an OpenAI-compatible interface while adding load balancing, guardrails, and cost tracking.
agentskills.io cybersecurity skills taxonomy for AI agents
agentskills.io provides a structured set of 817 cybersecurity skills for AI agents, mapped to multiple security frameworks and usable across popular agent coding tools.
Cognee AI memory platform for agents
Cognee is an open-source, self-hosted agent memory system that provides persistent long-term memory via a knowledge-graph engine across sessions.
OpenMontage
OpenMontage is an open-source, agentic video production system that wires many tools and pipelines into a programmable studio workflow.
How this page is made
A small daily pipeline, six steps.
Scroll down — the camera will tour each step against the live architecture diagram in the background.
Step · 01 · SOURCES
Four feeds.
GitHub Trending (HTML scrape, AI/LLM/agent keyword filter), Hacker News (Algolia API, last 24h), Medium (six RSS feeds: TDS, Gradient Ascent, AI Mind, The Generator, plus #llm and #large-language-models tags), and vendor blogs from Anthropic, OpenAI, DeepMind, Hugging Face. Each adapter is async and isolated — one source failing doesn't fail the run.
Step · 02 · SCRAPE
Fan out, merge in.
All four adapters fan out in parallel. Each respects per-source caps so one chatty feed can't drown out the others. The merged pool typically lands around ~70 candidates per run; on a slow news day it might be 20.
Step · 03 · DEDUP
Drop duplicates.
Two layers. First: within-batch clustering by cosine similarity ≥ 0.85 on text-embedding-3-small over title+excerpt — the same Anthropic announcement posted by four sources collapses to one candidate (longest excerpt wins). Second: backward dedup against the last 14 days of articles, by exact source_url and case-insensitive title.
Step · 04 · CURATE
Pick one winner.
An LLM ranks the survivors against the portfolio context (who reads this feed, what stack, what they care about) and picks one winner plus up to two alternates. The alternates are insurance — if the writer or critic later rejects the winner, the pipeline swaps in alternate #1 instead of dying.
Step · 05 · WRITE
Draft the post.
Same LLM, different prompt. The writer drafts a 200-450 word first-person post covering three beats: WHAT shipped, WHERE/WHEN it's useful (one concrete situation), WHAT TO LOOK AT FIRST (entry point, file, page, flag). One short bullet list is fine. No headings, no images, no marketing voice.
Step · 06 · CRITIC
Cheap quality gate.
A third LLM call, deliberately lenient — it only rejects empty/garbage drafts and only revises drafts >600 words, <80 words, or with stray headings/images. One revise attempt allowed; if the revised draft still fails, the pipeline swaps to the next alternate from the curate step instead of looping forever.
Step · 07 · PUBLISH
Land on this page.
Survivor gets persisted to two Postgres tables: articles (the story you just read up top) and pipeline_runs (the audit trail — scraped count, unique count, winner id, error if any). The Next.js page revalidates via ISR; the next visitor inside an hour gets the new article straight from the static cache.
The loop — once a day
Read what survived.
Seven steps, one cron, one article per day at 06:30 Europe/Warsaw via APScheduler. Scroll back up to today's pick — it came out the end of this pipeline this morning. Tomorrow's run is already scheduled.
Want this built for you?
This page is what an agent-curated internal-news system looks like at the smallest possible scale. Bigger ones are similar — same shape, more inputs, tighter rubric. If you want one for your team, talk to the lab.
Bio & channelsTalk to the lab→