Visual Analytics

Agent Usage & Telemetry

Aggregated efficiency, tool calls, and token metrics across autonomous AI coding sessions. Adhering to the data exploration mantra: Overview first, filter, then details on demand.

🔍 What is this data?

This dashboard reflects real tool execution telemetry and token usage collected directly from my local developer workstations (via harnez). Every time an orchestrated AI agent (such as Claude Code, OpenAI Codex, or Antigravity) reads files, executes shell commands, runs test suites, or applies diffs while pairing with me, the event is benchmarked for duration, success/failure, and quality rating.

Total Tool Invocations

Recorded tool executions

Overall Success Rate

Executions with exit code 0

Average Tool Score

1 to 5 quality rating scale

Distillation Savings

Noise stripped before context

Activity Timeline

Tool execution volume aggregated by day across active agent workflows.

Tool Breakdown & Performance

Details on demand: frequency, failure rates, duration, and human/automated quality scores per tool.

Tool Name	Calls	Failure Rate	Avg Score	Avg Duration	Raw Output Volume

Activity Classification & Work Breakdown

Taxonomic categorization of agent tasks (testing, building, refactoring, debugging, git). Zero-prose classified.

Activity Category	Calls	Share	Failures	Failure Rate

Agent Overview

Invocations and aggregate ratings by harness.

Agent	Invocations	Failures	Avg Score

Privacy & Sanitization Policy: All telemetry shown on this page is pre-aggregated and scrubbed via harnez usage export. Absolute filesystem paths, confidential system directories, private repositories, and sensitive notes are strictly stripped.