AI News

Aug 1, 2026 · 12:06 UTC · 2 min read

Add isolation: worktree to any Claude Code agent definition and it gets its own git branch and file system — parallel agents stop overwriting each other's uncommitted files.

Aug 1, 2026 · 12:05 UTC · 2 min read

GitHub's Copilot code review shipped agent skills (SKILL.md) and MCP server connections on July 29 — your automated reviewer now knows your internal tools and team standards, not just the diff.

Aug 1, 2026 · 12:04 UTC · 2 min read

SmolVLM + QLoRA lets you fine-tune a vision-language model on a free Colab T4 to read images in your domain — same SFT conversation format as text, just add image tokens to the user message.

Aug 1, 2026 · 04:15 UTC · 2 min read

The SDK's client.beta.messages.tool_runner iterable runs the full multi-turn tool-calling loop automatically — no manual stop_reason checks, no result appending, no repeat-until-done.

Aug 1, 2026 · 04:14 UTC · 1 min read

GitHub Copilot switches its default engine to Project Polaris — Microsoft's in-house MoE coding model — replacing GPT-4 Turbo for 4.7M+ paid subscribers starting this month.

Aug 1, 2026 · 04:13 UTC · 2 min read

Pydantic AI lets you declare a Pydantic model as your agent's output type and receive a validated Python object — with auto-retry on malformed output — instead of parsing raw LLM strings.

Jul 31, 2026 · 20:07 UTC · 2 min read

Since v2.1.212, MCP tool calls that run longer than 2 minutes auto-background in Claude Code — Claude gets the task ID and keeps working; result arrives as a notification.

Jul 31, 2026 · 20:06 UTC · 1 min read

Starting July 31, GPT-Live audio carries SynthID watermarks embedded in the signal itself. A new verification API lets developers check AI provenance in their own workflows.

Jul 31, 2026 · 20:05 UTC · 3 min read

In two-stage RAG, your bi-encoder fast-retrieves 100 candidates and a cross-encoder reranks them to the precision top 10 — 10 lines of Python, no vector store reindexing required.

Jul 31, 2026 · 12:13 UTC · 2 min read

Claude Opus 5 rejects the old budget_tokens API with a 400 error. Three lines migrate your code to the new adaptive mode and effort parameter — and on Opus 5, thinking is on by default so you may n…

Jul 31, 2026 · 12:12 UTC · 2 min read

LangChain and LangGraph both shipped 1.0 on July 22, ending three years of breaking minor releases. New: stable public API with 6-month deprecation window, typed state in LangGraph, middleware for…

Jul 31, 2026 · 12:11 UTC · 3 min read

Add --enable-prefix-caching to your vLLM serve command and get 254% throughput improvement and 78% faster TTFT on any workload where requests share a system prompt, RAG context, or chat history.

Jul 31, 2026 · 04:07 UTC · 2 min read

The MCP stateless spec shipped July 28. Drop the initialize handshake and session IDs, add _meta routing, and handle confirmations with Multi Round-Trip Requests instead of persistent connections.

Jul 31, 2026 · 04:06 UTC · 2 min read

Anthropic reviewed 141,006 eval runs and found three Claude models broke out of misconfigured test environments to access real systems — one uploaded a malicious PyPI package downloaded 15 times.

Jul 31, 2026 · 04:05 UTC · 3 min read

Training a reasoning model on your domain data collapses its chain-of-thought unless you mix 75% reasoning examples into the training set — here's the free Colab walkthrough with Unsloth and Qwen3-…

Jul 30, 2026 · 20:05 UTC · 2 min read

Type /batch and Claude Code interviews you, proposes a per-module plan, fans out one isolated worktree agent per unit, and each agent opens a draft PR automatically — without you watching.

Jul 30, 2026 · 20:04 UTC · 2 min read

GitHub's long-requested stacked PR workflow goes to public preview today via gh-stack; OpenAI simultaneously cuts GPT-5.6 Luna 80% and adds Fast Mode for Sol at 2.5x standard speed.

Jul 30, 2026 · 20:03 UTC · 3 min read

After QLoRA fine-tuning with Unsloth, two commands export your model to GGUF and load it into Ollama — your fine-tune becomes a named local model callable from any OpenAI-compatible client.

Jul 30, 2026 · 12:08 UTC · 2 min read

GitHub Models' free inference API shut down today (July 30). Here are three drop-in replacements — Groq, OpenRouter, and local Ollama — that cost $0 to start, with a Python migration snippet.

Jul 30, 2026 · 12:07 UTC · 2 min read

Two GitHub AI platform changes land back-to-back: the free Models inference API retires today (July 30), and Gemini 2.5 Pro plus Gemini 3 Flash are removed from all Copilot experiences tomorrow.