Skills go cross‑agent: install once, use in Claude, Cursor, Codex — plus July updates and autonomy trade‑offs

This week’s momentum in coding agents isn’t about one hero model — it’s about the glue. Skills are getting portable, orchestration tools are maturing, and even trace formats are landing so we can finally debug the agent as a system, not a vibe. Meanwhile, the autonomy vs. suggestion split is getting clearer, with Google’s internal Smith project clarifying the high‑autonomy tier even if you can’t use it yet.^[2]

The skills layer is standardizing across agents

DataRobot’s open skills pack is a clean example: ACP skill definitions that plug into Claude Code, Cursor, OpenAI Codex, Gemini CLI, and VS Code Copilot — with a universal installer so you can add everything or just one capability per agent.^[3] Even better, it embraces the different instruction formats each vendor already uses (Claude “Skills,” OpenAI AGENTS.md, Gemini extensions), so you don’t have to hand‑port the same integration five ways.^[3]

Install once, use everywhere (Claude, Cursor, Codex, Gemini CLI, Copilot)

Universal install (all skills for all supported agents):

npx ai-agent-skills install datarobot-oss/datarobot-agent-skills

Install a specific skill (example: predictions):

npx ai-agent-skills install datarobot-oss/datarobot-agent-skills/skills/datarobot-predictions

Target a specific agent (example: Cursor, then Claude):

npx ai-agent-skills install datarobot-oss/datarobot-agent-skills --agent cursor
npx ai-agent-skills install datarobot-oss/datarobot-agent-skills --agent claude

Claude Code also supports one‑line installs via the plugin marketplace:

claude plugin install datarobot-agent-skills@claude-plugins-official

Or from inside a Claude Code CLI session:

/plugin install datarobot-agent-skills@claude-plugins-official
/plugin marketplace add datarobot-oss/datarobot-agent-skills

All of the above comes straight from the DataRobot skills repo, and it’s designed for enterprise workflows like training, deployment, predictions, feature engineering, and monitoring — portable across major coding agents.^[3]

If you just want an index of agent skills across ecosystems, the community‑maintained Awesome‑Agent‑Skills list is a helpful map of what’s out there for Claude, OpenAI, Gemini, Copilot/VS Code, Cursor, and more.^[1]

Orchestrating teams vs in‑editor copilots

Two tracks are diverging:

• Persistent multi‑agent teams that carry objectives and memory across tasks. Example: agx advertises a team that coordinates work with 160+ merged PRs by the agents themselves — signaling the “autonomous crew” direction.^[1]
• In‑editor copilots focused on tight, low‑latency suggestions and chat. GitHub Copilot and Gemini Code Assist are the canonical baseline for this mode — fast suggestions, minimal ceremony, no background autonomy.^[2]

Google’s internal Smith AI sits firmly in the high‑autonomy camp (background execution, planning, multi‑step tasks, tight chat integration) but remains unreleased; today your practical high‑autonomy options are Claude Code, Devin, Cursor, or Google Jules, while suggestion‑first options remain Copilot, Gemini Code Assist, and ChatGPT/Codex.^[2]

Observability: trace your sessions

We finally have a lightweight, repo‑agnostic way to capture human‑agent sessions. The minitrace project proposes a session trace format for coding interactions, targeting observability and safety across frameworks. If you’re evaluating agents or trying to harden them, you want this kind of trace to ask, “What exactly happened between the suggestion, the tool call, and my broken test?”^[1]

This week’s versions: check and upgrade

If you haven’t updated lately, several CLIs ticked forward mid‑July. Havoptic’s tracker shows fresh releases for Gemini CLI (v0.51.0 on July 16), OpenAI Codex CLI (rust‑v0.144.5), GitHub Copilot CLI (v1.0.71), Antigravity CLI (v1.1.3), and Kiro CLI (v2.12.3). It also hosts compare pages for Claude Code vs these tools — handy when you’re deciding what to trial next.^[5]

Quick hygiene check (typical flags):

gemini --version
codex --version
copilot --version
antigravity --version
kiro --version

Choosing your agent today: autonomy vs suggestions

Developer chatter this month highlights the same rubric I use with teams: pick the tool that ships your feature fastest, not the one that tops a benchmark. Sentiment favors agents that can understand large repos, plan multi‑file changes, refactor safely, and stick with a long task — which is why Claude Code’s name keeps surfacing in discussions alongside Cursor, Gemini, DeepSeek, and others.^[4] If you want more autonomy, look at Claude Code, Devin, Cursor, or Jules; if you want low‑friction suggestions inside the editor, Copilot, Gemini Code Assist, and ChatGPT/Codex remain excellent.^[2]

Quickstart: wire a DataRobot skill into Claude Code and Cursor

Here’s a minimal, reproducible way to trial a cross‑agent skill in two environments the same afternoon.

1. Install the skill pack globally, then pin per agent:

# all skills, globally
npx ai-agent-skills install datarobot-oss/datarobot-agent-skills

# ensure Cursor picks it up
npx ai-agent-skills install datarobot-oss/datarobot-agent-skills --agent cursor

# ensure Claude Code picks it up
npx ai-agent-skills install datarobot-oss/datarobot-agent-skills --agent claude

2. Alternatively, install a single capability (predictions) if you want a narrower test:

npx ai-agent-skills install datarobot-oss/datarobot-agent-skills/skills/datarobot-predictions

3. For Claude Code specifically, verify via marketplace commands in the CLI session:

/plugin install datarobot-agent-skills@claude-plugins-official
/plugin marketplace add datarobot-oss/datarobot-agent-skills

These commands and compatibility notes come directly from the DataRobot skills documentation and work across major coding agents thanks to ACP skill definitions.^[3]

Key takeaways

• Skills are going portable: one definition, many agents. Start with the DataRobot ACP skills to feel the difference.^[3]
• Orchestration is real: try a persistent team (e.g., agx) if you need multi‑step autonomy; otherwise stick with in‑editor copilots for speed.^[1][2]
• Trace everything: adopt a session trace like minitrace to debug agent behavior over time.^[1]
• Update your CLIs: mid‑July releases hit Gemini, Codex, Copilot, Antigravity, and Kiro; check versions and retest flows.^[5]
• Choose by workflow, not hype: autonomy (Claude Code, Devin, Cursor, Jules) vs suggestions (Copilot, Gemini Code Assist, ChatGPT/Codex).^[2][4]

References

1. coding-assistants — https://github.com/topics/coding-assistants
2. Google Agent Smith AI: What It Is & How This Coding Agent Works — https://www.upgrad.com/blog/google-agent-smith-ai
3. GitHub – datarobot-oss/datarobot-agent-skills: Bring DataRobot platform capabilities to your coding agents — https://github.com/datarobot-oss/datarobot-agent-skills
4. The AI coding race is moving faster than ever. Just a year ago … — https://www.instagram.com/reel/Da4wSXQSbra
5. Claude Code v2.1.208 | Havoptic — https://www.havoptic.com/r/claude-code-2.1.208