Users report GPT-6 Sol weaker than 5.6 Sol on complex tasks

Claude Code v2.1.290 adds serverToolUses to a mod's turn.step hook result, exposing the API-run tool calls (the advisor) with id, name, input, and start/end times. Plugin hooks gain agentId on tool.check so a hook can distinguish a subagent's permission check from the main session's.

Browserbase's Miguel González Fernández and Microsoft Research's Corby Rosset present research on verifiers for computer-use and web agents, where an official judge scored agent success at 74% while a better verifier put it at 38%.

Amazon Bedrock adds GLM 5.3, an open-weight model aimed at coding and agentic workloads such as refactoring repositories spanning hundreds of files and sustaining multi-hour agentic workflows with tool use.

An APK teardown found a "Gemini Calling" intro screen with examples like "Call Mom and tell her I will be 15 minutes late" and "Call John and ask if he is coming for dinner." The teardown also surfaced granular per-app permissions, including phone call settings, though Google may never ship the expansion.

Laurie Voss of Arize AI demos a toy-store agent that fails a request for toys under $6 because its search tool lacks a price filter, then adds observability to trace the failure and has a coding agent patch it.