
llama.cpp 1.0 Ships: Stable API, FP8 Cache, Server Lockdown
llama.cpp 1.0.0 ships semver, FP8 KV cache, a host-memory MoE cache and server lockdown, but removes 35 deprecated API functions that can break upgrades.
Agents, coding tools, and local AI for builders
Hands-on coverage of AI for software developers: coding agents, Claude Code and Cursor, the Model Context Protocol, local and open-source models, and agentic workflows — tested and explained, not hyped.

llama.cpp 1.0.0 ships semver, FP8 KV cache, a host-memory MoE cache and server lockdown, but removes 35 deprecated API functions that can break upgrades.

Codex CLI 0.169.0 vs 0.168.0: codex exec untrusted directories stop loading repo MCP servers. Here is the CI migration path and the verdict.

A leaked Fireworks key led to a backdoored Claude Code extension in Cursor. Privacy Mode held; here is who was exposed and what to check.

Gemini 4 Hydrogen vs GPT-6.1 Terra: Google wins eight of nine benchmarks, but costs 50% more per token. Prices, scores and a routing rule.

Cursor Projects multiplayer lets teammates' agents coordinate, while Antigravity Swarm serves one developer. Our verdict for teams and solo users.

DeepSeek Harness ties for first in all five Harness Leaderboard evaluations, but one cleanup habit costs it Terminal-Bench. Verdict and setup.

Claude Code 2.1.323 adds experimental Code Mode. Our verdict: upgrade for the Bash permission fixes, and trial Code Mode only in a sandbox.

Shadcn improve passed 10,423 stars on GitHub Trending. We compare its plan-first audit skill with a manual Claude Code subagent split and pick a winner.

Claude Code 2.1.321 makes Fable 5.2 the default at $8/$40 per million tokens, 20% below Fable 5.1, with four sandbox and permission fixes.

Claude Code secrets security setup for small teams: Bitwarden Teams beats 1Password on cost for 2-3 people. A 30-minute audit plan.

Cursor Live Apps lets cloud agents build and host web apps with Postgres and SSO. Verdict: good for internal tools, not yet for customer-facing apps.

PLaMo 3 Translate open weights from Preferred Networks run offline in 53 languages. We compare it with TranslateGemma and Hy-MT2 and pick a winner.

The Chainalysis breach exposed password hashes and live tokens for all users. Here is how it compares with ransomware and what defenders should rotate.

AI news today for 9 November 2026: Claude Code's don't-ask-again rule leaks real secrets into public repos, and it outranks Anthropic's turbine order.

Arete 27B cuts Qwen3.8-27B's propaganda rate from 22% to 0.9% and runs free on a 24GB GPU. Benchmarks, Ollama setup and costs in INR.

OpenMausBot is the best open source AI agent platform for coding in 2026, running Claude, Codex or Grok locally under Apache 2.0.

Claude Code subagents now write status lines with Haiku 5.5, cutting per-update cost about 99%. Session savings are closer to 5%. Here is who gains.

Tencent Hy4 vs Mistral Large 4: Mistral ships downloadable weights and lower prices today, while Hy4 is API-only until Tencent publishes weights.

Claude Code message threads cut median time to first token about 35% on 500k-1M token requests. Setup steps, a comparison with caching, and limits.

Antigravity cheminformatics vs El Agente Fármaco: Google's skills win for teams that need a working agent now; the open preprint suits researchers.

Agentic AI vs generative AI examples with worked token math: when a one-shot draft is cheaper than an agent loop, priced in USD and INR.

Claude Sonnet price vs GLM-5.3 compared on real API rates, with monthly cost maths, an INR table and the cases where Sonnet still earns its rate.

The best AI trading bot setup in 2026 pairs ChatGPT with MCP backtesting tools, but the tested pro process still wins on real money. Verdict inside.

Agentic AI interview questions in 2026 test production operation: evals, containment and failure traces beat framework recitation.