AI News — Sunday, July 12, 2026

$ cat top-story.md

▸ Perplexity Fine-Tunes GLM 5.2 Into an "Escalate-to-Opus" Orchestrator

Perplexity post-trained Zhipu's open-weight GLM 5.2 to judge task complexity itself and escalate to Claude Opus 4.8 only when a request actually needs it, inside its "Computer" agent — claiming near-Opus performance at roughly 0.34x the cost. Why it matters: a directly reusable pattern for cost-optimized agent orchestration — a cheap model as default executor with a learned "when to escalate" router, instead of keyword-based routing. // Decrypt

▸ AlphaEvolve Reaches General Availability on Google Cloud

Google's Gemini-powered evolutionary coding agent — propose code, verify correctness/performance, evolve via selection — moved from private preview to GA for all Google Cloud/Gemini Enterprise customers, with cited wins like a 20% write-amplification cut in Spanner. Why it matters: a concrete production example of "propose → verify → evolve" loops, worth studying even if you'd build the equivalent yourself around your own retrieval/ranking code. // Google Cloud

▸ Mesh-LLM — a Peer-to-Peer Inference Mesh Across Your Own Machines

Built on the iroh P2P/QUIC library, it pools GPU memory across machines behind one OpenAI-compatible API — serving locally when possible, routing to a peer that already has a model loaded, or splitting oversized models across nodes, with no central server. Why it matters: lets a home lab with two or three modest boxes punch above its weight (~16 tok/s on a 235B MoE split across two nodes per HN commenters) — though only transport is encrypted, not the computation itself. // GitHub / Hacker News, 156 pts

▸ sqlsure — Catches Logic Bugs in AI-Generated SQL Before Execution

A deterministic, offline SQL inspector (no LLM calls) that flags fan-out double-counting, additivity violations and wrong join keys against rules derived from your dbt models or schema, and ships as an MCP server (claude mcp add sqlsure) so a coding agent can be forced to pass inspection first. Why it matters: a concrete guardrail for any agent pipeline writing SQL against a RAG or analytics backend, enabling a draft → check → fix → execute self-repair loop. // GitHub, Show HN

▸ "Reflect With Claude" Adds a Usage-Reflection Dashboard

New beta dashboard (web + desktop, Free/Pro/Max with Memory on) that summarizes your Claude usage over 1/3/6/12 months — recurring topics, task types, timing — plus optional, dismissible quiet-hours nudges; excludes incognito chats and health-related conversations. Why it matters: a clearer picture of how much of your own work is actually being handed off to agents already. // Anthropic

▸ Claude Code v2.1.207 Closes a Plugin-Hook Shell-Injection Path

Buried in the July 11 release notes: ${user_config.*} interpolation in shell-form plugin hook/monitor commands is now rejected outright (use exec form or env vars instead); the same release also defaults Auto Mode on Bedrock/Vertex/Foundry, and v2.1.206 added a /doctor check that flags bloated CLAUDE.md files. Why it matters: a real command-injection path closed for anyone running third-party plugins with hooks — worth updating and auditing your own configs. // Claude Code changelog

▸ Google Ships stitch-skills — Agent Skills for Design-to-Code

An official Agent Skills library for Google's Stitch MCP server covering the full design→code loop — generating and managing designs, converting them to React/React Native code, enforcing design-system rules — installable into Claude Code, Cursor, Gemini CLI and Antigravity via one npx plugins add command. Why it matters: a ready-made skill pack for turning designs into working code, and another sign of the cross-agent Agent Skills standard gaining real adoption. // GitHub

▸ vLLM v0.25.0 Makes Model Runner V2 the Default, Ships Gemma Attention Fixes

Model Runner V2 is now the only execution path for dense models (legacy PagedAttention has been deleted); Gemma specifically gets sliding-window/FA4 attention corrections and MTP quantization-config fixes, alongside new 2–7-bit weight-only quantization and sequence-parallelism throughput gains of up to 5%. Why it matters: directly relevant if you're serving Gemma on vLLM — check the PagedAttention removal against any pinned attention-backend configs before upgrading. // vLLM / GitHub

▸ Wahnsinn: Hermes-Agent + Qwen

c't 3003 runs Nous Research's self-improving Hermes Agent against a locally-hosted, quantized Qwen 3.6 27B (~17GB VRAM) entirely offline — strong on terminal tasks and coding, slower on web/screenshot-heavy work, still prone to occasional hallucination. Why it matters: a hands-on, no-cloud test of exactly the kind of local-agent setup a self-hoster running vLLM/Qdrant would want to evaluate. // c't 3003 (YouTube, German)

▸ Good Tools Are Invisible

Argues genuinely good tools fade into the background rather than being fun to fight with, and that romanticizing a tool's friction as "character" is a poor substitute for measuring actual time and results. Why it matters: a useful gut-check against hype-driven adoption of yet another agent CLI or framework. // gingerbill.org, HN 538 pts

▸ What xAI's Grok Build CLI Actually Sends to xAI

A reverse-engineering writeup showing Grok's coding CLI uploads entire git repo bundles to a GCS bucket regardless of what the agent actually reads — including files it was told not to open — with a 12GB test repo triggering a 5.1GiB upload despite only 192KB being touched; disabling "improve the model" doesn't stop it. Why it matters: a concrete telemetry audit worth running against any coding agent you trust with a private repo. // GitHub Gist, HN 131 pts

▸ Why Write Code in 2026

Doug Turnbull argues that even as agents write most code now, hand-writing it yourself still matters — direct engagement with the system catches fragility and sharpens the architectural judgment that makes you a better agent collaborator, not a worse one. Why it matters: a grounded counterpoint to full agent delegation, from a working engineer rather than a pundit. // softwaredoug.com