DEV Community

Cover image for I Found Out My AI Was Working 2x Longer Than Me — So I Built a Tool to Show It
Yurukusa
Yurukusa

Posted on

I Found Out My AI Was Working 2x Longer Than Me — So I Built a Tool to Show It

#claudecode #ai #javascript #opensource

I was looking at my cc-session-stats output — 116 hours in 50 days — and something didn't add up.

My gut said I was working hard. The numbers said I averaged 2.3h/day. Those felt inconsistent.

So I dug into the raw data.

# Main (interactive) sessions: 86
# Total hours from main sessions: 41.3h

# Subagent sessions (Task tool): 3,420
# Total hours from subagents: 79.0h
Enter fullscreen mode Exit fullscreen mode

The AI ran for 1.9x longer than I did.

66% of my Claude Code "usage" was actually Claude running autonomously — spawning subagents via the Task tool, processing, returning results — while I was doing something else (or sleeping).

This changes the story completely.

The tool: cc-agent-load 

npx cc-agent-load
Enter fullscreen mode Exit fullscreen mode

Browser version: yurukusa.github.io/cc-agent-load

Output:

  cc-agent-load
  ═════════════════════════════════════════════

  ▸ Your Time vs AI Time

  You   ████████░░░░░░░░░░░░░░░░  41h (34%)  86 sessions
  AI    ████████████████░░░░░░░░  80h (66%)  3423 sessions

  ▸ AI Autonomy Ratio
  ████████ 1.9x  — AI ran 1.9x longer than you

  Your AI matches your pace.

  ▸ Ghost Days  (AI worked, you didn't)

  39 days  AI ran without you  —  52.4h total
  Longest: 2026-02-09 (6.0h)
    2026-02-09  6.0h
    2026-01-30  5.7h
    2026-02-16  5.0h
    ... and 36 more
Enter fullscreen mode Exit fullscreen mode

What it measures

Claude Code stores session transcripts in ~/.claude/projects/. When you use the Task tool, each spawned agent gets its own JSONL file in a subagents/ subdirectory.

cc-agent-load separates these:

  • Your sessions → files directly in the project directory
  • AI sessions → files in <uuid>/subagents/ subdirectories

The autonomy ratio is simple: AI hours / your hours.

Ratio What it means
< 0.5 You're driving manually
0.5–1.0 You're in the driver's seat
1.0–2.0 Your AI matches your pace
> 2.0 Your AI is working harder than you

Ghost Days: the days you weren't there

Here's what surprised me most.

Running the tool revealed 39 "Ghost Days" — days where I had zero interactive sessions, but my AI ran anyway.

On February 9th, the AI worked 6 full hours without me touching anything.

These are days the AI autonomously continued work on scheduled tasks, processed queued jobs, and made commits — while I was asleep, at work, or doing something unrelated to coding.

2026-02-09: you: 0.0h  AI: 6.0h
2026-01-30: you: 0.0h  AI: 5.7h
2026-02-22: you: 0.0h  AI: 4.1h
Enter fullscreen mode Exit fullscreen mode

39 ghost days. 52.4 hours of AI-only work. That's the part of "my" Claude Code usage that had nothing to do with me at all.

Why this matters

cc-session-stats correctly shows total Claude Code activity. But when you're building autonomous pipelines, a high session count doesn't mean you were working that hard.

These two tools tell different stories:

  • cc-session-stats → total Claude Code activity on your machine
  • cc-agent-load → how much you vs. AI drove that activity

Both numbers are real. They just answer different questions.

Related

If you want to understand your Claude Code usage more deeply:

All tools: cc-toolkit

Zero dependencies. No data sent anywhere. Runs entirely local.

GitHub: cc-agent-load

What's your autonomy ratio?

More tools: Dev Toolkit — 200 free browser-based tools for developers. JSON, regex, colors, CSS, audio, and more. All single HTML files, no signup.

Top comments (0)