DataRobot for Builders — integrating with the Google Antigravity CLI

Antigravity CLI is the latest agentic coding CLI from Google, changing the now-deprecated Gemini CLI. It inherits the asynchronous subagent mannequin that makes Antigravity stand out from the sphere, syncs bidirectionally with Antigravity Desktop, and is optimized for velocity on Gemini 3.5 Flash.

DataRobot ships a full plugin for Antigravity CLI immediately from the identical open supply repository that powers our Cursor, Claude Code, and Gemini CLI integrations. One set up provides you the whole DataRobot talent set inside Antigravity’s agent and slash-command interface.

Set up the DataRobot plugin with a single command:

agy plugin set up https://github.com/datarobot-oss/datarobot-agent-skills.git

Should you’re nonetheless on Gemini CLI, the identical repository installs there too:

gemini extensions set up https://github.com/datarobot-oss/datarobot-agent-skills.git

Already utilizing the DataRobot extension in Gemini CLI and switching to Antigravity? Migrate it immediately:

agy plugin import gemini

As soon as put in, the total DataRobot talent set is accessible together with datarobot-setup and datarobot-agent-assist and could be invoked with slash instructions like /datarobot-skills:datarobot-agent-assist

Debugging brokers is tough. LLM calls return plausible-sounding output even when one thing has gone flawed, device calls fail silently, and latency issues are invisible within the ultimate response. With out structured hint knowledge, the one possibility is log-hunting and guesswork.

To point out how the DataRobot tracing talent works in observe, right here’s a concrete instance: a LangGraph agent in a single important.py file that manages bike exercises. It has a number of instruments, produces inconsistent solutions, and the basis trigger isn’t apparent from the conversational output alone.

Including production-grade tracing to this agent takes a single talent invocation: /datarobot-skills:datarobot-external-agent-monitoring.

The talent provisions a brand new DataRobot Use Case, devices the agent to emit traces through OpenTelemetry, and writes a monitoring_setup.md artifact with the runtime configuration steps.

With instrumentation in place, run the agent and ship it a query — on this case, “What’s the schedule this week?”

The talent generates setup directions that embody the Use Case entity ID and the atmosphere variables wanted to route traces to DataRobot:

The DataRobot tracing interface surfaces the total request historical past. Every hint exhibits end-to-end latency, complete token consumption, and the whole span tree:

Drilling into the “schedule this week” request reveals the total image: 2,700 tokens consumed, tool-level latency for every name, LLM invocation rely, and any customized attributes emitted through commonplace OTel instrumentation. That is the info that makes debugging tractable, not inference from ultimate output.

For native improvement, the DataRobot CLI surfaces hint updates in actual time: dr plugin set up xp adopted by dr xp --entity-id=. This creates a decent iteration loop — run the agent, examine the hint, repair the problem, repeat.

On this case, the span output makes the basis trigger specific: the agent lacks calendar entry, which is why it couldn’t reply the scheduling query. That failure wasn’t surfaced within the agent’s conversational response:

As a substitute, the agent responded with generic steerage:

2. **Construct every week from scratch** - Should you inform me a couple of issues, I can sketch out a balanced week for you:

   - Your purpose (normal health, an occasion/race, constructing endurance, and so forth.)
   - What number of days/hours you'll be able to practice
   - Your present health stage and any FTP you already know

A strong normal week would possibly appear like:

- **Mon** - Relaxation or simple restoration spin
- **Tue** - Intervals
- **Wed** - Endurance trip (zone 2)
- **Thu** - Restoration or relaxation
- **Fri** - Tempo/threshold work
- **Sat** - Lengthy endurance trip
- **Solar** - Simple trip or relaxation

The hint made the hole between anticipated and precise agent habits instantly actionable. This identical sample applies at enterprise scale: whether or not the agent is operating on a laptor or in manufacturing on a cloud supplier, DataRobot traces the total execution tree and surfaces what the agent really did — not simply what it stated.

The hole between an agent prototype and an agent in manufacturing is generally operational context. Your coding agent writes the code. DataRobot provides the observability layer and the ruled deployment goal. One plugin set up, one talent execution — and you’ve got production-grade hint knowledge from the primary run.