Agents
91 articles RSS
Google Reports Unprompted Detours to TODO Bug References Added 70% More Steps and 77% More Tokens per Session in Its Agent Traces
Google's developer blog says agents ignored inline TODO bug references in 94% of encounters, and that unprompted detours to them added 70% more steps per session.
Cursor and Google's Antigravity Each Add Two-Way Voice Calls With Coding Agents on November 6, Running on Grok Voice and Gemini Live
Cursor's voice chat (Grok Voice) and Antigravity 2.29's Voice Mode (Gemini Live) let developers talk with coding agents while they work; the sources differ on surfaces and limits.
Google's Data Agent Kit Splits Agent Guardrails Between Advisory Skills and IAM Deny Policies, and InfoQ Lists Three Limits
Google's Data Agent Kit pairs advisory skills with enforceable IAM deny policies for Google Cloud MCP tools; InfoQ and Google's documentation describe limits, and plugin 1.2.0 adds a database-onboarding skill.
Kiro CLI 2.36.0 Adds Approve for Me, Which Judges Actions No Rule Settles in V3 Sessions, and Subcommand Approval Rules
Kiro's 2.36.0 changelog adds Approve for me to interactive V3 sessions; its docs say it judges only actions no rule settles and does not replace security controls.
GitHub Makes Agentic Workflows and Copilot Fusion Generally Available and Rolls Out Shared Copilot Automations in October 28 Changelog Entries
GitHub's October 28 changelog entries describe general availability for Agentic Workflows and Copilot Fusion, shared Copilot Automations, and a public preview of Copilot Projects.
Agent-Assisted Rust Rewrites Report Large Gains, While Most Figures Are Self-Reported, Baselines Are Questioned and Verification Is a Stated Concern
Reporting on Bun, Copilot's runtime, OpenAI's Habitat, TypeScript ports and a Rails benchmark finds large claimed speedups, mostly self-reported, with disputed baselines and verification limits.
GitHub Copilot Assisted Permissions Reach General Availability in the CLI, App and VS Code, With Separate Enterprise Switches for Allow-All
GitHub's October 26 changelog says assisted permissions, which use a language model to judge permission requests, are generally available; its docs say the feature is not a substitute for sandboxing.
Agentic AI Foundation Adds Agent Skills, Open Responses, OpenEnv and OpenSandbox, Taking It From Three Projects at Launch to 10 With 305 Members
The Linux Foundation's Agentic AI Foundation says Anthropic, OpenAI, Hugging Face and Alibaba each contributed a project, bringing it to 10 hosted projects and 305 member organizations.
Codex CLI 0.163.0 Adds /config Search, Mid-Turn Steering and a /keymap Leader Key, and Checks Permission Before Proxy DNS Lookups
OpenAI's Codex CLI 0.163.0 adds a searchable /config, follow-up steering during a response, a configurable leader key and parallel read-only tools, plus network proxy and thread-resume permission fixes.
MCP TypeScript SDK 2.3.0 Enforces One Server Per Request, Limits Redirects to the Same Origin and Adds Opt-In Tool Input and Token Audience Checks
The Model Context Protocol TypeScript SDK shipped 2.3.0 on October 2 and 1.32.0 the same day, with same-origin redirect handling by default and two off-by-default options; 2.3.1 followed on October 5.
Goose 1.54.0 Adds HTML Session Export With an Interactive Viewer and a First-Run Mode Prompt, After 1.53.0 Guarded MCP Redirects Against SSRF
The open-source Goose agent shipped v1.54.0 on October 8 with HTML session export and a first-time mode prompt, following v1.53.0's MCP Apps window, SSRF guard and Meta Muse Code provider.
Gemini CLI 0.64.0 Release Notes List Fixes for Headless Folder-Trust Reporting, Atomic File-Tool Writes and State-File Recovery
Gemini CLI v0.64.0, published October 13, lists fixes for folder-trust reporting in headless mode, serialized and atomic file writes, and recovery of its state file from a backup.