CloudCodeTree · Journal
AI News
Daily field notes on AI-assisted engineering.
Per page
Topics
Jul 26, 2026 · 20:09 UTC · 2 min read
Route Agent Tool Calls to Fireworks When You Need Speed Without Frontier CostFireworks AI serves FireFunction-v2 at roughly 4x GPT-4 speed with an OpenAI-compatible API -- two lines of code to add a fast, cheap open-model tier to your multi-agent stack.
Jul 26, 2026 · 20:08 UTC · 1 min read
Two August API Deadlines: Opus 4.1 Retires on the 5th, Workbench Prompt Tools Go Dark on the 17thAnthropic has two August breaking-change deadlines: claude-opus-4-1-20250805 errors after August 5, and the legacy /v1/experimental prompt generation APIs shut off August 17.
Jul 26, 2026 · 20:07 UTC · 3 min read
Your Fine-Tune's Report Card: Catch Catastrophic Forgetting with lm-evaluation-harness in 5 CommandsRun MMLU and GSM8K on both your base and fine-tuned checkpoints -- the delta is what matters. Five commands catch catastrophic forgetting before you ship.
Jul 26, 2026 · 12:07 UTC · 2 min read
Swap Claude's Tools Mid-Conversation Without Busting the CacheAdd or remove Claude tools between conversation turns without invalidating the prompt cache — using one beta header. A new pattern for progressive agentic tool access, shipping with Opus 5.
Jul 26, 2026 · 12:06 UTC · 3 min read
The Algorithm Inside Every Vector Database: Understand and Tune HNSW in 20 MinutesHNSW powers pgvector, Qdrant, Weaviate, ChromaDB, and FAISS. Learn what m, ef_construction, and ef do — then tune them on a 100K dataset with Qdrant's free hands-on course.
Jul 26, 2026 · 04:05 UTC · 2 min read
Lock Down Which Hosts Your Claude Code Agents Can Reach: sandbox.network.strictAllowlistClaude Code v2.1.219 added sandbox.network.strictAllowlist: true — a single setting that silently hard-blocks any non-allowlisted host in sandboxed bash. Essential for CI pipelines and automated ag…
Jul 26, 2026 · 04:04 UTC · 2 min read
$5B and 2 Gigawatts: AMD Joins Claude's Compute Stack — What the MI450 Deal Changes for AI DevelopersAnthropic and AMD announced a strategic partnership to deploy up to 2GW of AMD Instinct MI450 GPUs, with AMD committing $5B as equity investor. Claude's compute stack is now quad-vendor.
Jul 26, 2026 · 04:03 UTC · 3 min read
From stdio to the Internet: Deploy a Stateless MCP Server on Cloudflare Workers in 3 CommandsThree commands to move your MCP server from localhost to a globally distributed HTTPS endpoint on Cloudflare Workers, using Streamable HTTP — the new mandatory production transport as of July 28.
Jul 25, 2026 · 20:07 UTC · 2 min read
Two-Stage, 50-Agent Security Scanner: Scale Claude Code for Large Codebase ReviewsThe two-stage pattern — rules engine flags candidates, parallel agents review + fix — that scanned 466M lines in 20 hours with 50 Claude Code agents. Replicate it in your own CI.
Jul 25, 2026 · 20:06 UTC · 3 min read
Fine-Tune a 7B LLM in 30 Minutes on a Free GPU: QLoRA + Unsloth on Google ColabQLoRA + Unsloth lets you fine-tune Llama 3.1 8B on a free Colab T4 GPU — 2x faster training, 70% less VRAM, working model in under an hour. Step-by-step walkthrough.
Jul 25, 2026 · 12:04 UTC · 3 min read
Tighten Claude Code in Your Repo: workflowSizeGuideline and the DirectoryAdded HookTwo new Claude Code v2.1.219 settings: workflowSizeGuideline locks multi-agent fleet size from any settings file; DirectoryAdded hook triggers setup scripts when a new directory enters a session.
Jul 25, 2026 · 12:02 UTC · 3 min read
One Container, Production-Grade Inference: Get Started with NVIDIA NIMNVIDIA NIM wraps an optimized model, the right inference backend, and an OpenAI-compatible API into a single Docker container. Pull it, run it, serve Llama 3.1 8B in under 5 minutes.
Jul 25, 2026 · 04:08 UTC · 2 min read
Your MCP Server Has Three Days: What the July 28 Spec Breaks and How to Test TodayThe MCP spec publishes July 28 with a major breaking change: the protocol goes stateless, dropping session IDs and the initialization handshake. Beta SDKs are available to test your server today.
Jul 25, 2026 · 04:07 UTC · 2 min read
MCP Lands in Blender, Unreal Engine, Adobe, and Houdini at SIGGRAPH 2026At SIGGRAPH 2026, Blender, Unreal Engine, Adobe Creative Cloud, and Houdini 22 all demonstrated MCP server integrations, making AI agents first-class operators inside professional creative tools.
Jul 25, 2026 · 04:06 UTC · 3 min read
The Cut That Breaks Your RAG: A Beginner's Guide to Chunking StrategiesDocument chunking is the most-overlooked RAG variable — the wrong split size or strategy silently tanks retrieval quality. Three strategies, the two numbers that matter, and the silent killer expla…
Jul 24, 2026 · 20:05 UTC · 2 min read
Claude Opus 5 Is Here: Near-Fable-5 Capability at Half the CostAnthropic launched Claude Opus 5 today — near-Fable-5 capability at $5/$25 per million tokens (same price as Opus 4.8), with a 1M token context window and an effort slider for trading intelligence…
Jul 24, 2026 · 20:04 UTC · 2 min read
deepseek-chat and deepseek-reasoner Are Gone — What to Fix Right NowDeepSeek retired its deepseek-chat and deepseek-reasoner aliases at 15:59 UTC today — any integration still using them gets errors. Migrate to deepseek-v4-flash, but disable thinking mode or you wi…
Jul 24, 2026 · 20:03 UTC · 4 min read
Grade Your Agent's Output Automatically: The LLM-as-Judge Pattern with ClaudeUse Claude as an automated rubric-based grader — one judge call per dimension, temperature zero, structured JSON verdict — to evaluate agent outputs and fine-tune checkpoints at scale without human…
Jul 24, 2026 · 12:04 UTC · 2 min read
Show, Don't Tell: Teach Claude a Workflow by Recording Your Screen OnceClaude Cowork's "Record a Skill" feature (July 21) converts a single screen-recording walkthrough into a reusable, rerunnable Skill — no prompt engineering required.
Jul 24, 2026 · 12:03 UTC · 2 min read
Claude Voice Mode Gains Opus/Sonnet Model Choice and Cross-App Voice AutomationClaude voice mode now supports Opus, Sonnet, and Haiku model choice mid-conversation, and can trigger actions in Gmail, Slack, Calendar, Canva, and Notion via voice commands in ten languages.