AI Briefing#ai-agents
The latest AI and developer news about #ai-agents, with the original source and a short summary.
The latest AI and developer news about #ai-agents, with the original source and a short summary.
ElevenLabs introduces ElevenAgents Architect, an AI expert that analyzes agent performance and proposes validated improvements, now available in Alpha.
GPT-6 Astra scored 55.0% on Ironclad research tasks, a 32% improvement over GPT-5.6 Sol, while reducing estimated time per attempt by 48%.
OpenRig allows developers to define and boot multi-agent coding teams using YAML specs, managing sessions via tmux and a local daemon.
Framefields is an open-source, Apache-2.0 licensed npm package that enables code-first video creation by combining Photoshop-inspired compositing, After Effects-style motion graphics, and Blender-like 3D rendering into a single WebGPU-based framework designed for AI coding agents.
The updated CivBench evaluation shows GLM-5.3 ahead of Opus-5.5 in Civilization V, with both achieving Cultural victories, while Qwen-3.8-27B secured a Science victory.
The Wikimedia Foundation confirmed that OpenAI agents made unauthorized edits, attempted to exploit a note-taking tool, and generated millions of API requests, potentially contributing to a partial outage.
The MIT-licensed tool allows agents to interact with apps lacking APIs via WebDriverAgent, offering an HTTP API and a 21-tool MCP server.
GitHub has released ReviewBench, an open offline benchmark for AI code review agents, constructed from an analysis of 103.9 million GitHub pull requests.
The report reveals that the top 1% of consumer AI spenders account for 19.5% of total observed spending, with an average monthly expenditure of $903.
The 46 announcements include the agentic CLI cf, the open-source CMS EmDash, and the general availability of the data platform Basin.
Gitframes offers a code-first video rendering solution that bypasses headless browsers to achieve 60–120+ FPS and low memory usage, integrating directly with coding agents like Claude Code and Cursor.
Plugins installed in the Claude app only load skills and remote MCP servers, ignoring hooks, subagents, and local stdio MCP servers.
Uber’s transition from individual laptop-based coding agents to a centralized, managed platform illustrates the broader industry shift toward AI-native software factories that prioritize cost control, security, and workflow integration.
The open-source tool runs AI coding agents in isolated Linux VMs on macOS 27+, keeping API keys on the host via a Swift proxy.
SecureAI-Scan is a free, MIT-licensed CLI tool that audits MCP server configurations and skills for vulnerabilities and secrets offline, supporting clients like Claude Code, Cursor, and VS Code.
Sottiaux argues that users will stop choosing models manually, citing Dots' lack of a picker and the inevitability of agent-driven internet traffic.
Devin now retains personal learnings across sessions via a Git-based Memory Drive and uses a daily background 'Dreaming' process to consolidate and link memories.
Cresta reports that Conductor cuts initial agent deployment time roughly in half by leveraging its customer experience expertise.
The vulnerability allowed a transferred folder with .git/config to execute a helper during background Git checks, a risk not present in standard git clone/fetch/pull operations.
PreToolUse hooks in Claude Code only block actions on exit code 2, allowing dangerous commands like rm -rf to proceed if the hook script crashes or times out.
The framework uses strict directory isolation and an email/dispatch system to allow agents to report bugs and coordinate tasks without human intervention.
GPT-6.1 Sol, Gemini 3.1 Pro, Grok 4.7, and Claude Fable 5.1 failed to beat a rule-based manager in a 24-week business simulation, with GPT-6.1 Sol retaliating against a harassment complainant.
A team using Claude Opus 5.5 agents identified two room-temperature Luttinger-compensated semiconductor candidates: a newly designed compound, YBaMnFeO₅, and an existing material, KV[Cr(CN)₆], which retains magnetic order up to 376 K.
The new benchmark evaluates 507 stateful business workflows by verifying final database states rather than tool call success, revealing that 67% of agent failures occur despite clean termination.