AI Agents: Every NEW Claude & Codex Feature Explained

TL;DR

• 1

  Codex /goal changes how long-running agent work gets done — Riley shows OpenAI’s new /goal command framing the outcome instead of a single prompt, with examples of people reporting runs lasting 4 hours and even 1 day 14 hours.

• 2

  Anthropic’s talent grab got much louder with Andrej Karpathy — Riley calls Karpathy joining Anthropic a huge signal, alongside CTO-level defections from companies like Supabase, Adept, You.com, and Workday, all to become individual contributors.

• 3

  Claude now has a multitasking terminal mode, but product fragmentation still bugs him — the new claude agents interface lets him fire off five parallel research tasks in the terminal, yet he still thinks the split between Claude Code and Claude’s co-work/desktop experience is one of Anthropic’s biggest mistakes.

• 4

  Codex is turning into a real super app, not just a coding tool — plugin sharing across teams, annotation-based design mode, and the command-command “appshots” shortcut let Codex understand any app on-screen, take screenshots, and even type directly into Google Docs using computer use.

• 5

  Google’s Gemini Spark looks promising, but Riley says Google still lacks a clear center of gravity — Spark appears to blend chat, cloud agents, folders, Google app connections, Canvas, Deep Research, NotebookLM, and media tools, but he says Google is spread across Gemini, AI Studio, and “anti-gravity” with no single flagship workflow.

• 6

  Cursor is emerging as the dark horse super app contender — Riley says Cursor’s new Composer 2.5 model is extremely fast, cheap enough that “an hour” might cost under $1, and paired with an in-app browser, integrations, and automations that make it look increasingly like Codex with momentum.