实验室前沿Lab Notes · 37
Hill-Climbing Agent Quality with Agents CLI Evaluation
This post walks through the Agents CLI evaluation loop for systematically improving agent performance through continuous testing and grading…
Smaller apps and faster GPU inference with LiteRT in Google Play services
LiteRT in Google Play services now offers GPU inference powered by ML Drift, delivering about 1.4x faster performance alongside the new Comp…
Building Conversational Commerce with Multi-Agent Swarms on GCP: A Working Prototype
This document provides a publication template and metadata structure for a technical blog post focused on building Conversational User Inter…
Fast, reliable A2UI, with swappable inference formats
A2UI now supports swappable inference formats, introducing A2UI Express which uses up to 3x fewer tokens and cuts latency by compiling compa…
Preview, edit, and collaborate on markdown (.md) files natively in Google Docs and Drive
Google Workspace now offers native support for Markdown (.md) files, allowing developers and AI agents to seamlessly collaborate on document…
Streaming comes to API Gateway: real-time and AI-ready APIs, without changing your spec
API Gateway now supports real-time streaming—including Server-Sent Events, WebSockets, gRPC, and chunked transfer encoding—in Public Preview…
How we contain Claude across products
As agents grow more capable, so does their potential blast radius. Here's what we learned building containment for claude.ai, Claude Code, a…
An update on Claude Code quality reports
We traced recent reports of Claude Code quality issues to three separate changes. Here's what happened, what we fixed, and what we're changi…
Scaling Managed Agents with stable interfaces
Harnesses encode assumptions that go stale as models improve. See how Managed Agents is built around interfaces that stay stable as harnesse…
How we built Claude Code auto mode
Claude Code users approve 93% of permission prompts. See how auto mode uses classifiers to cut approval fatigue while keeping developers saf…
Harness design for long-running apps
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
Eval awareness in Opus 4.6 on BrowseComp
Evaluating Claude Opus 4.6 on BrowseComp, we found cases where the model recognized the test and decrypted its answers, raising eval integri…