This is the news. Agent dossiers — when to pick one, when to skip it — are on the agents page. Practical guides support your own testing and deployment decisions.
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.
AWQ 4-bit quantization shoves a 70B model onto a single 24GB GPU with negligible quality loss. vLLM's Marlin kernel serves AWQ weights at near-FP16 throughput. Here is the production-ready setup with a working OpenAI-compatible endpoint, a load test, and the three flags that decide whether it actually runs.
Every agent framework is a loop. Models stop, you run tools, you feed results back, repeat. Here is the whole agent runtime in ~40 lines of plain Python against the Anthropic API.
Stop losing Claude Code context every time you close your laptop. A five-minute tmux setup keeps the agent, the scrollback, and the conversation alive through sleep, SSH drops, and lunch breaks. One pattern, every project, zero extra runtime.
A 60-line Redis wrapper that gives you exact-match dedup, prefix-based mass-invalidation on prompt changes, sliding expiry on hot keys, and stale-while-revalidate — the version of LLM caching you reach for when one worker becomes a fleet.
Bi-encoder retrieval finds the right neighborhood. A cross-encoder reranker picks the actual answer. The two-stage pattern takes 30 lines, runs in 50ms, and adds 15-25% recall on real RAG workloads.
Your coding agent works one branch at a time. You can run five of them in five directories on five branches from a single repo, with no stashing, no rebasing, and no merge conflicts at submission time. Git worktrees have shipped in core git for a decade — this is the workflow I use to parallelize Claude Code, Codex, and Aider on the same codebase.
You are re-running the same 200 prompts every time you tweak a temperature. Stop paying for it. Twenty lines of Python, a SQLite file, and your dev eval loop goes from $3.40 per run to $0.00. Here is the recipe.
Claude Code reads one file at the repo root: CLAUDE.md. Most teams never write one, and that is why it keeps guessing wrong. Twelve lines fixes it.
Stop paying the model-load tax on every prompt edit. Bun's built-in --hot flag plus a long-lived Ollama client gives you sub-second prompt iteration. Sixty seconds of setup, real afternoon-long productivity.
Everyone's building AI agents. Nobody wants to admit they're just expensive middleware with a personality disorder and delusions of agency.
Every major LLM now has a 'give me valid JSON' mode. They're not created equal. A practical breakdown of how Claude, GPT, and Gemini handle structured output — with real code and the gotchas nobody puts in the docs.
StepFun shipped a 198B sparse-MoE vision-language model that takes #1 on ClawEval-1.1 (67.1), costs $0.20/M input, runs on a DGX Spark and a 128GB Mac Studio, and is open-weights on Hugging Face. Yes, really.
Qwen3.7-Plus on Bailian June 2: multimodal, 1M context, five-step agentic loop, $0.40/$1.60 per 1M token (60% cheaper than text-only Max). The story: Alibaba closed the open-weights moat and the price cut is the tell.
You push, GitHub Actions spins, a bot takes four minutes, and half the comments are noise. Stop. A 50-line pre-commit hook reviews your staged diff with Claude before the push. Five minutes from zero to a working reviewer that blocks on real bugs and costs $0.005 per commit.
There's a quiet revolution happening in JavaScript tooling, and it runs on Rust. OXC — the JavaScript Oxidation Compiler — is building a suite of high-performance tools that make your current build chain look embarrassingly slow. The numbers are absurd. Here's why it matters.
Half the RAG stacks I have torn down this year solve a problem the underlying model stopped having in late 2025. With 1M-2M token context windows now default, retrieval is a tax on latency and an excuse to skip real evaluation.