CloudCodeTree · Journal
AI News
Daily field notes on AI-assisted engineering.
cloudcodetree.com/feed.xml · full text, no tracking
Per page
Topics
Aug 28, 2026 · 04:06 UTC · 2 min read
Auto-Continue When Your Claude Code Session Hits the Usage Limit: The Setting That Saves Overnight RunsClaude Code v2.1.234 auto-continues your session when a claude.ai usage limit resets — unattended sessions no longer stall indefinitely when they hit the cap. On by default in /config.
Aug 28, 2026 · 04:05 UTC · 2 min read
Claude Cowork Gets a Built-In Browser; GPT-5.6 Sol Dropped 20% a Week AgoClaude Cowork now opens a side-panel browser for web tasks — no Playwright required. GPT-5.6 Sol API input/output prices dropped over 20% on August 21.
Aug 28, 2026 · 04:04 UTC · 2 min read
Run Any Open-Weight Model Locally in Two Commands: Ollama and Its OpenAI-Compatible APIPull and run Qwen, Llama, Gemma, or DeepSeek locally with one command and call it from Python using the exact same OpenAI API code — just change one URL. No cloud cost, no data leaving your machine.
Aug 27, 2026 · 20:11 UTC · 2 min read
Your Claude Code Sessions Can Now Coordinate: A Practical Guide to Agent TeamsClaude Code's experimental agent teams (one flag away) let a lead session spawn independent teammates from a shared task list — parallel implementation, research, or review without manual coordinat…
Aug 27, 2026 · 20:10 UTC · 2 min read
The Research Bench Is Reshuffling: Google DeepMind's Brain Drain in NumbersDeepMind's share of top European AI hires fell from 49% to 18.6%. John Jumper (AlphaFold) joined Anthropic, Noam Shazeer (Gemini co-lead) joined OpenAI — data from Fortune, August 27.
Aug 27, 2026 · 20:09 UTC · 3 min read
You Trained the Adapter — Now Prove It Worked: A Three-Step Eval for Fine-Tuned LLMsTraining loss dropping isn't proof of improvement. Three checks — held-out perplexity, lighteval benchmark, and a task-specific metric — tell you whether your LoRA adapter actually helped.
Aug 27, 2026 · 12:06 UTC · 2 min read
Batch Your Offline Claude Calls and Pay Half: The Message Batches APIThe Anthropic Message Batches API runs up to 100K async requests at 50% of synchronous pricing. Right for eval runs, nightly reports, and bulk tagging; wrong for anything real-time.
Aug 27, 2026 · 12:05 UTC · 3 min read
From Documents to Answers in 30 Lines: Build a Minimal RAG Pipeline With LangChainBuild a working RAG pipeline in 30 lines: load documents, split into chunks, embed with a free local model, store in Chroma, retrieve by similarity, and generate with Claude. Three failure modes ex…
Aug 27, 2026 · 04:03 UTC · 2 min read
Every Claude API Response Now Carries an Invisible Watermark — What to Adjust Before You ShipNew Claude models embed an invisible text watermark and C2PA metadata on file outputs, globally since August 2. Presence proves Claude was involved; absence proves nothing. Here is what to adjust.
Aug 27, 2026 · 04:02 UTC · 2 min read
OpenAI's Custom Inference Chip Outperforms NVIDIA Blackwell on Throughput per Watt — and GPT-5.6 Luna Falls 80%OpenAI published Jalapeño benchmark results: 1.5–1.9× more throughput per watt vs. current NVIDIA Blackwell. Same week, GPT-5.6 Luna dropped 80% to $0.20/1M input tokens.
Aug 27, 2026 · 04:01 UTC · 3 min read
Turn Any Python Function Into a Tool Claude Code Can Call: Your First MCP Server in 15 LinesThe MCP Python SDK turns decorated Python functions into tools Claude Code can call. Install mcp[cli], decorate your functions, test with mcp dev, wire to Claude Code in settings.json.
Aug 26, 2026 · 20:04 UTC · 3 min read
Give Your Managed Agent a Strategist to Consult Mid-Turn: The Advisor Model PatternManaged Agents lets you pair a fast executor model with a more capable advisor it can consult mid-turn — paying Opus-level tokens only for the hard decisions, not the whole run.
Aug 26, 2026 · 20:03 UTC · 1 min read
Manage Your Claude Organization Without curl: Admin API Ships to Seven SDKsAnthropic's Admin API — org members, workspaces, API keys, rate limits, service accounts, CMEK — is now accessible in Python, TypeScript, Go, Java, Ruby, PHP, and C# SDKs under client.beta.organiza…
Aug 26, 2026 · 12:07 UTC · 3 min read
Your API-Key Claude Code Session Loses Its Cache After Five Minutes — Set This to Survive Long BreaksAPI-key and cloud-provider users get a 5-minute prompt cache by default; one setting in settings.json extends it to one hour and eliminates full-recompute costs when you step away from a long session.
Aug 26, 2026 · 12:06 UTC · 3 min read
Instruction-Tune a Llama 8B on Your Own Data for Free: QLoRA and Unsloth on Colab's T4 GPUQLoRA 4-bit quantization cuts Llama 3.1 8B from ~16 GB to ~5 GB VRAM; Unsloth's LoRA adapters then train on Colab's free T4 in about 30 minutes. You get a portable adapter to export as GGUF or push…
Aug 26, 2026 · 04:08 UTC · 2 min read
Two Managed Settings Keep Your Whole Team on the Approved Model List in Claude CodeClaude Code v2.1.243 adds modelPicker (curate the /model list to approved options) and modelPricing (plug in contracted rates for accurate cost estimates). Solves the "four developers on four diffe…
Aug 26, 2026 · 04:07 UTC · 2 min read
Langflow Has an RCE on CISA's Exploit List; LangGraph Has a CVE Chain — Patch Your Agent InfrastructureLangflow has a CISA-listed RCE under active exploitation since before its July 17 patch; LangGraph has a three-CVE chain that exposes your LLM API keys. Both have patches — check your versions.
Aug 26, 2026 · 04:06 UTC · 3 min read
Your Retriever Picks the Wrong Chunk Half the Time — Combine BM25 and Dense Vectors with RRFHybrid search combines BM25 keyword retrieval with dense vector search, fused via Reciprocal Rank Fusion — no score normalization needed. Outperforms either alone on most real corpora.
Aug 25, 2026 · 12:08 UTC · 3 min read
Stop Rule-Listing: The Six Context Shifts That Cut Claude Code's Internal Prompt by 80%Anthropic removed 80%+ of Claude Code's system prompt for Claude 5 models with no quality loss. Six shifts — rules to judgment, upfront to progressive disclosure — apply directly to your CLAUDE.md.
Aug 25, 2026 · 12:07 UTC · 3 min read
Q4_K_M or AWQ? The Quantization Decision That Fits a 70B Model Into Your Existing GPU BudgetQuantization lets a 70B model at 4-bit outperform a 7B at full precision. Here is the format decision: GGUF Q4_K_M for Ollama and local inference, AWQ for NVIDIA production serving with vLLM.