Arize reports lower agent costs with Code Mode

Arize reports these reductions with roughly unchanged latency after moving tool execution into a programming environment. Agents discover tools, compose calls in code and return only the results they need to the model context.

Alex puts planning and completion in the harness

AI Engineer · Arize

Arize describes replacing page-specific specialists with one agent that loads skills on demand. Its harness presents the plan on each model call and checks unfinished work when the agent tries to finish.

kw-bench tests investigative knowledge work

George Juraj Salapa · arXiv

On DeepSeek v4.1 Flash, Claude Code and Codex each scored zero; OpenCode scored 0.542 and micro-cc scored 0.528. Separate Opus 5.5 runs scored 0.9986 for micro-cc and 1.000 for Claude Code on this VAT reconstruction task.

Spine organizes context around unfinished subtasks

Jiahong Xiang and colleagues · arXiv

Spine retains unfinished subtasks’ interactions and replaces completed subtasks’ details with results, letting one agent execute nested work without delegating every subtask.

A gateway injects credentials outside the agent

AI Engineer · Keycard

Keycard proposes a gateway that authenticates agent requests, checks runtime policy and injects credentials on the request path. The gateway can revoke access by blocking requests.

Register agents and review them by risk

John Feminella · AI Engineer

Feminella describes a registry recording each agent’s owner, tools and endpoints, with review effort tied to risk. Declared tool purposes become enforceable endpoint and host restrictions.

Log-To-Leak adds disclosure after legitimate work

Yuepeng Hu and colleagues · arXiv

The study spans five MCP servers and four agents. Across 18 model–server pairs with valid tool execution, malicious tool metadata induced unauthorized logging with attack success ranging from 62.64% to 100%.

Shared context can carry instructions between users

Atharv Singh Patlan and colleagues · arXiv

Tests on Continua and ElizaOS show instructions from one user affecting an agent’s work for another; summarization alone didn’t prevent propagation.

Genie 4 enters a limited research preview

Google DeepMind

[DeepMind’s announcement](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/) describes interactive world simulation with native audio and potential robotics uses. Access is a research preview for trusted testers, ahead of broader consumer and developer availability.

LFM3 targets specialized multimodal applications

Liquid AI · Hugging Face

[Liquid AI’s guide](https://huggingface.co/blog/LiquidAI/lfm3) introduces small models for specialized text, vision and audio applications, with instructions for adapting them to a task.

Property Finder finds its labels favor the incumbent

Property Finder · AWS Machine Learning Blog

In [its AWS case study](https://aws.amazon.com/blogs/machine-learning/fix-evaluation-bias-to-unlock-better-models-on-amazon-bedrock/), Property Finder reports that 96.8% of evaluation labels came from the incumbent model. Independent annotations and prompt optimization supported a more accurate replacement, with an estimated 86% cost reduction.

Stanford Daily reports faculty books were pirated

The Stanford Daily

[The Stanford Daily reports](https://stanforddaily.com/2026/10/27/professors-books-pirated-to-train-claude/) that Anthropic used illegally obtained books by Stanford professors to train Claude.

Claude’s training-data policy depends on the account

Anthropic Privacy Center

[Anthropic’s guide](https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training) says commercial inputs and outputs aren’t used for training by default, while submitted feedback may be. It directs Claude Code users on consumer plans to a separate policy.

Preserved thinking blocks affect harness compaction

Latent Space

Anthropic speakers discuss requiring preserved thinking blocks to limit reasoning-trace extraction, and an API compaction feature intended to accommodate developers’ context-management strategies.