52 articles · 5 categories
The finishable daily brief
Thursday, Nov 12, 2026
52 articles · 5 categories
read top to bottom · then stop
In 30 seconds
Verifying agent-written code was today's main thread. A new hook catches coding agents that weaken tests to turn CI green, Martin Fowler's site argues customer tests are now the human-AI contract, and one essay says human review cannot keep up with agent-built inference engines.
Anthropic found that groups do worse when every user brings their own agent. On the serving side, Google open-sourced XSE and vLLM shipped day-0 support for Inkling 2 Small and MiniMax H3-Image.
Prefer to watch?
Five pieces treat verification, not generation, as the bottleneck once agents write both the code and its tests.
A deterministic, offline hook for Claude Code, Codex and Cursor that diffs each change for loosened assertions, skipped or deleted tests, lowered coverage thresholds and CI steps allowed to fail.
Argues that passing latency and throughput benchmarks leaves rare concurrency bugs hidden in agent-written inference engines, and that human review cannot scale to 1,000 parallel coding agents.
Martin Fowler's site argues that when AI writes both production and test code, customer tests rather than developer tests become the contract between humans and the agent.
An agent-added retry can double-charge a customer when a timeout hides a successful charge: the feature works, the test suite grows, and the system gets worse.
Across 324,088 commits in 52 top AI-tooling repos, the median repo has an AI signature on 16.9% of commits from its last 180 active days; 10 repos are at 50% or more.
Anthropic measured a multi-agent coordination failure, while four open-source projects put controls at the sandbox, the tool call, the MCP config and the prompt.
In four simulated environments with models from Anthropic, OpenAI and Alibaba, group outcomes were worse when each user delegated to their own agent than when one agent served everyone.
A zero-trust sandbox runtime that claims 10x the density of standard containers, resume in under 500ms and more than 500 activations per second, with isolated compute, identity, networking and storage per agent.
Checks every agent tool call against plain-English rules before it runs, and screens each tool result for prompt injection before the agent reads it.
Inventories an organization's agents and MCP servers, works out what each can read, write, execute and send, and tests that access against 12 controls mapped to NIST 800-53 and SOC 2.
swapback is a fail-closed local proxy that swaps names, emails, phone numbers and keys for placeholders before a prompt leaves the machine, then restores them in replies, streams and tool calls.
Google open-sourced two pieces of its serving stack, and vLLM shipped same-day support for two new models with published throughput numbers.
Google open-sourced XSE, the XLA Serving Engine, which takes the Python interpreter out of the serving path and handles six million queries a second across Search, Ads and YouTube Shorts.
vLLM serves Thinking Machines' multimodal MoE model (594B total, 17B active parameters, up to 1M tokens of context) at up to 630 tokens per second per user, with LoRA support.
vLLM-Omni runs MiniMax H3-Image on four NVIDIA Vera Rubin GPUs at a 0.87-second median for native 4096×4096 images; a separate throughput configuration reaches 6,115 images per hour.
Google open-sourced TPU Sync for moving KV cache and model weights on TPUs.
AWS walks through serving frontier models on its Trainium chips with SGLang-Neuron.
Production models turn over in weeks, a hosted model API is going away, and one team found its discount math was wrong.
Vercel's November AI Gateway Production Index, drawn from more than 100 trillion tokens a month, puts the effective half-life of a model in production at about six weeks.
Batch is half price on every provider that offers it, yet it is the smallest of three cost levers, does not stack with caching the way it appears to, and loses to a tier drop.
A customer posted an email from Inception Labs saying the Mercury API that hosts its diffusion models is shutting down; no blog post had been published yet.
Databricks and a new embedded graph engine both redesign the data layer around agents as the caller.
Databricks targets the sustained concurrency of agents, which can run hundreds of queries per second where a human analyst runs a few dozen in an afternoon.
A Rust property-graph engine in a single binary that speaks Cypher 25 and stores a graph as one file, a directory or in memory, with Python, Node.js, WASM and Rust bindings.
Generally available: Genie answers in DMs and channels under the asker's Unity Catalog permissions, and updates documents and tickets through MCP connectors after approval.
You are caught up for this edition
