Production-tested skills for AI agents. Every skill is security-scanned, tier-rated, and verified. Browse by ecosystem or category below.
July 9, 2026: OpenAI moves GPT-5.6 to GA — three tiers (Sol, Terra, Luna), 1M context, Programmatic Tool Calling, Ultra mode. The interesting story isn't the benchmarks. It's the structure.
Claude Code will happily `rm -rf` the wrong directory. Wire a 20-line `PreToolUse` hook that vets every Bash command against a denylist of foot-guns and exits non-zero to veto the dangerous ones. Twenty lines, no seatbelt excuses.
Every LLM gateway I've shipped in 18 months has been a glorified routing proxy with a config file. Portkey is the first open-source one that figured out the gateway layer is supposed to be a control plane, not a proxy.
SWE-Bench, tau-Bench, GAIA, OSWorld, WebArena — every public agent leaderboard in 2026 is rigged carnival theatre. Labs know it. Your CFO does not. The score that matters is the one you run yourself on your own tickets.
Mistral shipped Leanstral 1.5 — an Apache-2.0 open-weights MoE (119B total, 6.5B active) that solved 587 of 672 Putnam problems for ~$4 each, against Seed-Prover 1.5's $300+. Formal verification just became cheap.
Pre-counting tokens stops the obvious cost incidents. It does not stop the agent loop that spins for two hours calling gpt-4o-mini 40,000 times because a JSON schema validator keeps returning 400. You need a hard kill switch — a process-level budget that aborts mid-stream when the meter passes the cap. Here is the build, 30 lines.
Every 'Prompt Engineer — $250K + equity' LinkedIn post is a self-own. The job description is shallow, the skill ceiling is depressing, the vendors are writing the role to shift blame off their models, and the tooling that replaces hand-tuned prompting is shipping every six weeks. Hire an AI engineer. Retire the title.
SpaceXAI shipped Grok 4.5 yesterday at $2 input / $6 output per million tokens — roughly 10x cheaper than Opus 4.8 per solved SWE-Bench task once you count the 4.2x token reduction. The real moat is the Cursor data flywheel, not the leaderboard.
Most LLM serving stacks bolt structured generation on top of a chat-completion engine. SGLang flips the script — a frontend DSL and runtime co-designed for branching, parallel, and constrained LLM programs. If your workload looks like an agent, a tree-of-thought, or a JSON-schema-heavy pipeline, it deserves a serious look.
Stand up a Postgres + pgvector backend that indexes millions of embeddings and answers nearest-neighbor queries in milliseconds — with the exact schema, index choice, and a working query loop.
Your fancy multi-agent orchestration is one LLM pretending to be a team. Stop shipping it. Stop selling it. Stop pretending the trench coat is full.
Anthropic shipped Claude Sonnet 5 on June 30 with near-Opus-4.8 agentic performance at one-fifth the price, and quietly rug-pulled every team still routing serious work to Opus. The tokenizer change is a 30% price hike nobody is talking about.
The Model Context Protocol went from Anthropic experiment to industry standard in six months. Here is what it actually gets right, where production implementations are still a mess, and what you should be doing today.
Five frameworks I deleted in 2026 are good software. The sixth is 137 lines of Python handling 11M turns a week. Here is exactly when to use which runtime — and the 20-line agent.py you should have shipped instead of your agent platform.
The demos are real. The production systems are mostly theater. Here is the hard technical map of what computer use agents can actually do in July 2026, the three architectural approaches, the six failure modes that kill deployments, and the production stack with code that works.
The skills economy is the new software supply chain. 180,000+ skills on the public registries, 1M+ in private use, growing 4,200/week on the public side and 25,000/week across enterprise marketplaces. The SKILL.md format is the universal agent skill artifact, the install path is one command, the four marketplaces that matter are Skills.sh, the Claude Skills Registry, Hugging Face Skills Hub, and Microsoft Copilot Studio. Here is the format, the install mechanics, the security model, the economics ($380M direct, $1.5B annualized), and the stack I'd ship Monday morning if I were a team building a real skill product this quarter.
Voice is the new default UI. The STT → LLM → TTS pipeline that 80% of teams are shipping is dead for production use. The right architecture in 2026 is end-to-end speech-to-speech with the model as the agent, and the four production systems I'd actually build on are OpenAI Realtime, Gemini Live, Kyutai Moshi, and Sesame CSM. Here is the full stack, the code, the latency benchmarks, and the cost numbers from four production deployments.
The chat box is the new front door. The browser is the new agent runtime. Perplexity Comet, ChatGPT Atlas, and Google's AI Mode have crossed 100M daily users between them, and the architectural pattern most teams are shipping (sidebar AI on top of Chrome) is wrong. Here is the engineering stack I would build if I were shipping a consumer AI browser in July 2026: Chromium fork with native MCP, Rust agent loop in the browser process, three-tier model strategy (local / mid / frontier), and on-device user-controlled memory.
WebGPU plus WebLLM plus transformers.js reached production in 2026. A 3B-parameter model now runs entirely in the browser, calls tools, and persists memory across sessions. The cloud is now the accelerator, not the runtime. Here is the full stack: 94 lines of TypeScript, a 4 GB model file, and zero API keys.
Open-source video models have been silent for two years. Skywork AI's SkyReels-V4 ships joint video and audio generation in a single forward pass at 1080p/32 FPS, and the dual-stream MMDiT architecture is a serious bet on independent unimodal objectives over the obvious cross-attention fusion everyone else tried first.
OpenAI dropped GPT-5.6 today — three tiers named Sol, Terra, and Luna at $5/$30, $2.50/$15, and $1/$6 per million tokens — and locked all of them behind a 'limited preview' for roughly 20 vetted partners at the U.S. government's request. The model is real, the benchmarks are real, and the rollout is the story.
Seven stories that defined the week: Anthropic's models pulled mid-launch, three frontier models dropped in two weeks, Google bet its enterprise future on agents, and the unit economics of premium AI subscriptions started looking like a problem nobody wants to name.
Pinecone, Weaviate, Qdrant, and the rest of the dedicated vector store layer are a 2022 bet that lost. Postgres swallowed them. The query you are paying $30k/month to run is a JOIN with an ORDER BY and a LIMIT.
Unsloth hit 2x faster training and 70% less VRAM by hand-writing Triton kernels for attention, RoPE, cross-entropy, RMS layernorm, MLP, and QKV — and then shipped a no-code web UI (Studio) in March 2026. With 500+ supported models, Apache 2.0, and a one-line GGUF export, it is the most disciplined open-source fine-tuning library shipping in 2026, and the boring infrastructure choice that just became the ambitious one.