How to use the Monitor Experiment Results agent skill

What the skill makes the agent do

The skill is a written workflow in a file called SKILL.md. When you ask the agent to check results, it follows these steps: identify where the experiment ran (local, SSH, Vast.ai or Modal), check whether the process is still running, capture the last lines of each screen session, look for JSON result files, and pull Weights & Biases metrics if the project uses them.

It then presents the raw numbers in a comparison table, flags anything unexpected such as NaN values, negative deltas or divergence, and suggests next steps. For Vast.ai and Modal jobs it also asks the agent to note instance state, hourly cost and cloud cost exposure.

• Checks running sessions with screen on the server
• Captures the last 50 lines of each screen session
• Lists and reads recent JSON result files
• Pulls W&B run history and links to run dashboards
• Can send an optional completion notification to Feishu

Install the monitor-experiment skill

Download the ZIP for this skill from our page. The skill package contains one file, SKILL.md. Unzip it so that the monitor-experiment folder ends up in one of these locations:

For all of your projects, unzip into ~/.claude/skills/. For a single project, unzip into that project's .claude/skills/ folder.

If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the ZIP there.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

How to trigger the skill with natural requests

You do not need to type the skill name. The skill description tells the agent to use it when you say things like "check results", "is it done", "monitor", or ask for experiment output. Say what you want in your own words and mention the backend if the agent might not know it.

The skill reads AGENTS.md, run notes or the launch summary to find the backend, so keeping those files up to date helps the agent pick the right place to look.

• "Check results for the sweep on my GPU server."
• "Is it done yet? The Vast.ai job started this morning."
• "Monitor the training run and tell me if the loss diverged."
• "Pull the W&B metrics for the last 20 runs and include the dashboard links."

Example: checking a remote screen session

For runs launched over SSH inside screen, the skill first lists the sessions, then copies the output of each one to a temporary file and reads the tail. The commands below are the ones written in the skill source. Replace the placeholders with your own server and session names.

If hardcopy fails, the skill says to check for log files or tee output instead.

What the skill needs before it can work

The skill can only monitor runs the agent can reach. For remote runs you need SSH access to the experiment server, and screen must be available on that server. If you use Weights & Biases, Python 3 with the wandb package and an API key are needed. If you launched jobs on Vast.ai or Modal, the agent needs access to your account for those services.

The Feishu notification is optional. It reads a config file at ~/.codex/feishu.json, and if that file is missing or its mode is set to "off", the step is skipped.

If the agent has no SSH or cloud credentials, the commands will fail and the skill cannot work around that.

• SSH access and screen on the remote server (remote runs)
• Python 3, the wandb package and an API key (if using W&B)
• Vast.ai or Modal account access (if those backends were used)
• Optional Feishu config at ~/.codex/feishu.json

Safety notes: scripts and network access

This skill runs scripts and uses the network. The commands it issues connect to your servers over SSH, query Weights & Biases through its Python API, and may read Vast.ai or Modal instance state and cost. If you do not want the agent querying billing or instance state, do not use this skill for cloud jobs.

Read SKILL.md before installing so you know exactly which commands the agent will run. The skill itself only reads results: the commands shown in its source fetch screen output, list files and read metrics.

• The skill runs shell commands over SSH and Python scripts that call the wandb API
• Cloud checks include cost exposure, which some users may not want queried
• Read the file before installing; the skill is released under the MIT licence

Tips for better monitoring results

Tell the agent which backend the run used, especially if it was Vast.ai or Modal, so it does not assume a plain SSH screen session. If W&B is configured but cannot be reached, the skill tells the agent to report the connection problem and fall back to screen, log and JSON evidence rather than treating the missing data as a failed experiment.

Ask the agent to compare against the correct baseline, meaning one with the same configuration. The skill's key rules say to show raw numbers before interpretation and to check training logs for errors before concluding that results are wrong.

• Name the backend (SSH, Vast.ai, Modal or local) in your request
• Keep the baseline config in your notes so comparisons are fair
• Ask for W&B dashboard links to share with reviewers
• Ask the agent to confirm whether runs are still going before reading final numbers