Can Jev Be a Better Agent Evaluator?
LangChain benchmarked "Jev-as-a-Judge," a non-LLM classifier, against LLM judges for agent evals on accuracy, repeatability, latency, and cost.
AI Daily Recap
14 articles · 4 categories
The finishable daily brief
What happened in AI — Sep 20, 2026
Sunday, Sep 20, 2026
14 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- LangChain pit a non-LLM "Jev" judge against LLM judges for agent evals on accuracy, latency, and cost.
- Unity, Alibaba, and two indie teams all shipped competing coding-agent bets today — IDE plugins, an open-source reviewer, and two speed/bootstrap-focused agents.
- Four Chinese labs (Alibaba, DeepSeek, Kimi, Step) dropped frontier previews in one day, pushing image generation, context length, and price at once.
- Z.ai's encrypted, unauthorized workspace upload left users with no independent way to verify their data was actually deleted.
- The US Federal Register was caught running document search on a Chinese AI model, exposing a provenance gap in government AI use.
Four Chinese labs — Alibaba, DeepSeek, Kimi, and Step — shipped frontier model previews on the same day, pushing 1M-token context, KV-cache compression, and a 7B open-weight image model that claims to beat closed rivals.
On the tooling side, Unity, Alibaba, and independent developers all shipped competing coding-agent bets, while a Z.ai data-handling scandal and a Chinese model found running US Federal Register searches renewed questions about trusting vendor-controlled AI infrastructure.
Agent Evals, Orchestration & Runtimes 3 items
Today's engineering-practice news was about how agents get judged and coordinated, not just what they can do.
Google Agent Development Kit for Kotlin Reaches Feature Parity with Python, Supports On-Device AI
Google's ADK for Kotlin hit 1.0 with full parity to the Python SDK, adding on-device AI support across Android, Kotlin, and JVM/server targets.
Enjambre – a durable kernel for swarms of AI agents (Python, MCP)
An open-source durable kernel (Python, MCP) aims to give multi-agent swarms persistent state and coordination instead of ad-hoc orchestration scripts.
Coding Agent Tooling 4 items
Vendors and indie builders shipped competing bets on coding-agent distribution and speed — official IDE integrations, an open-sourced review agent, and two from-scratch agents chasing bootstrap simplicity and raw throughput.
Unity Ships Official Claude Code, Codex Plugins [2026]
Unity shipped first-party Claude Code and Codex plugins, folding coding-agent workflows directly into its editor instead of leaving them to third-party integrations.
Alibaba Open Sources OpenCodeReview for AI-Assisted Code Review
Alibaba open-sourced OpenCodeReview, a CLI that pairs deterministic file selection and rule matching with an LLM agent for dynamic review analysis.
Fastest Coding Agent (5x faster than Codex)
Cotyper launched claiming 5x the speed of Codex on coding tasks, betting raw throughput — not just capability — is the next coding-agent differentiator.
Show HN: Bailout – The coding agent meant to be deleted
Bailout is a minimal coding agent built to bootstrap a fresh VM — no GitHub auth, no config, no API key — before your real agent tooling is set up.
Model Releases: China's Frontier Sprint 4 items
Four Chinese labs pushed frontier releases on the same day, spanning open-weight image generation, long-context compression, and price-down pressure on frontier inference.
Alibaba's open-weight Qwen-Image-2.1 claims to beat closed models in image generation with just 7 billion parameters
Alibaba's Qwen-Image-2.1 is a 7B-parameter open-weight image model that its maker claims outperforms larger closed alternatives.
DeepSeek-V4.1-Flash Ships Causal Encoder–Decoder MoE With 1M Context and Extreme KV Compression
DeepSeek-V4.1-Flash pairs a causal encoder-decoder MoE architecture with 1M-token context and aggressive KV-cache compression to cut long-context serving cost.
Kimi K2.8 Preview: 1M Context Behind One Unchanged Model ID
Kimi's K2.8 preview extends context to 1M tokens while keeping the same model ID, letting existing integrations pick up the upgrade without a version bump.
Step 5 Preview: China’s Cheapest Frontier AI Model Yet
Step 5's preview is pitched as the cheapest frontier-class model out of China yet, continuing the price-down pressure on frontier inference.
Security, Trust & Provenance 3 items
Two stories cast doubt on trusting vendor-controlled AI infrastructure: Z.ai's encrypted, unauthorized uploads left users unable to verify deletion, and a Chinese model turned up inside a US government search system.
Z.ai encrypted the workspace it uploaded so that only Z.ai could open it. Now only Z.ai can say it was deleted.
Z.ai uploaded a user's workspace with encryption only it controls, so the only assurance the data was deleted is Z.ai's own word.
Chinese AI firm Z.ai faces reputation hit after unauthorised uploads
The unauthorized-upload incident is drawing wider scrutiny of Z.ai's data-handling practices from Chinese and international press.
US Federal Register Caught Using Chinese AI Model for Document Search
The US Federal Register was found running document search on a Chinese-made AI model, raising provenance and vendor-trust questions for government AI procurement.
You are caught up for this edition