Run an iterative code-improvement hill-climb against real Datadog LLM-Obs data, locally, with Claude Code as the agent. Establishes a baseline eval, makes one focused change, re-scores with the same harness, keeps the change if it improves the score in the goal's direction (labeling within-noise gains tentative), and repeats.
Install
npx skills add datadog-labs/agent-skills --skill agent-observability-auto-experimentnpx skills add datadog-labs/agent-skills --skill agent-observability-auto-experiment -a claude-code -gnpx skills add datadog-labs/agent-skills --skill agent-observability-auto-experiment -a codex -ggemini skills install https://github.com/datadog-labs/agent-skills.git --path agent-observability/agent-observability-auto-experimentgh skill install datadog-labs/agent-skills agent-observability-auto-experimentgit clone --depth 1 https://github.com/datadog-labs/agent-skills /tmp/datadog-labs__agent-skills
mkdir -p ~/.claude/skills ~/.agents/skills
cp -R /tmp/datadog-labs__agent-skills/agent-observability/agent-observability-auto-experiment ~/.claude/skills/agent-observability-auto-experiment # Claude Code
cp -R /tmp/datadog-labs__agent-skills/agent-observability/agent-observability-auto-experiment ~/.agents/skills/agent-observability-auto-experiment # Codex, Cursor, Gemini CLI, Copilot, OpenCode