What's New
Game-changing features you should know about
Claude Code
Agent Teams — Peer-to-Peer AI Collaboration
Multiple Claude agents can now communicate directly with each other, sharing context and coordinating complex tasks. Think code review → fix → re-review cycles, all autonomous.
Hooks System — Event-Driven Automation
PreToolUse, PostToolUse, PreCompact, PostCompact hooks let you automate linting, security checks, and context preservation. Claude Code becomes programmable.
Background Agents + Git Worktrees
Spawn agents that work in isolated git worktrees in the background. Start a refactor, keep coding — the agent creates a branch with its changes when done.
GPT Codex
Layered AGENTS.md — Global to Local Repo Context
Codex now works best with an instruction chain: personal guidance, repo rules, and nested overrides. You can keep durable team conventions close to the code instead of repeating them in every task.
Sandbox + Approval Policies — Fast Local Automation with Guardrails
Local Codex is most practical when sandbox mode and approval policy are tuned together. Workspace-write plus on-request gives you fast edits and test runs without handing over your whole machine.
Cloud Environments — Setup Scripts, Offline Agents, Optional Internet
Cloud Codex separates setup from execution: install and bootstrap first, then run the agent in a tighter environment. Internet access stays off by default and can be narrowed with allowlists when needed.
Gemini CLI
Plan Mode — Read-Only Design Sandbox
Before editing any code, use /plan to enter a strict read-only mode to explore and draft your architecture. Only approved plans get executed.
Sub-agents — Delegate to Experts
Offload batch refactoring to the Generalist and bug hunting to the Codebase Investigator. Save your main context window and maximize throughput.
Agent Skills — Override Agent Behavior
Summon specialized workflows like security-review, or use skill-creator to teach the agent your team's unique deployment checklist.