Agent Observability

What This Skill Does

The Agent Observability skill provides a comprehensive monitoring stack for your OpenClaw agents. It is designed to bridge the visibility gap between high-level task execution and low-level agent performance. By installing this tool, you gain the ability to track productivity metrics, audit decision-making logic, capture trace data on failed subagent tasks, and monitor drift against your core INTENT.md instructions. The system consists of four primary components: a weekly throughput dashboard for high-level health checks, a decision audit log that records the 'why' behind agent choices, a failure tracer that acts as an automated 'black box' for subagent errors, and a drift-guard that ensures your autonomous agents remain aligned with their original objectives.

Installation

To install, run the following command: clawhub install openclaw/skills/skills/donovanpankratz-del/agent-observability. After installation, manually move the provided files into your local scripts/ and lib/ directories. You must schedule the throughput-dashboard.js and drift-guard-auto.js in your weekly cron jobs or heartbeat script. For deeper integration, import the audit and tracer libraries into your agent's core decision-making and validation loops as outlined in the implementation documentation.

Use Cases

This skill is ideal for teams deploying OpenClaw agents in production environments where auditing is mandatory. Use it to debug persistent subagent failures by examining generated JSON trace logs. It is also perfect for long-term project management, allowing developers to assess the economic cost and quality of agent-generated code over extended periods. Finally, it provides a 'fail-safe' mechanism to catch 'intent drift' where agents gradually deviate from project scope.

Example Prompts

1. "OpenClaw, run the throughput dashboard script and summarize my agent's productivity for this week."
2. "Show me the last five entries in the decision audit log to understand why the agent initiated a new file structure."
3. "Run a drift-guard check against my current INTENT.md file and report any significant deviations in agent behavior."

Tips & Limitations

Ensure that you manually wire the decision-audit function into high-stakes decision points, as the system cannot automatically guess the importance of every internal process. The failure-tracer relies on a quality score; if your agent-scoring mechanism is not calibrated correctly, the tracer may trigger excessively or not at all. Periodically purge the memory/traces/ directory to save disk space, as logs can grow rapidly during high-volume subagent tasks.