My AI agents burned tokens all night doing nothing

Something felt off when I checked my Claude Code usage this week. One day had 809 sessions. The usual is around 100.

I run OpenFang, an open-source agent OS, as a background daemon on my laptop. I’d activated the built-in Researcher Hand through the dashboard a few days ago, used it once, then forgot about it.

It didn’t forget about me.

The numbers

I dug into the session logs stored under ~/.claude/projects/. 497 session files for the openfang project, almost all from a single day.

After parsing through the JSONL, the picture was clear:

The 346 code reviewer sessions were from the superpowers plugin doing its thing during development. Fine. Expected.

The 170 Researcher Hand sessions were not expected.

Every 60 seconds, like clockwork

The timestamps told the story:

2026-03-19T13:06:57
2026-03-19T13:07:57
2026-03-19T13:08:57
2026-03-19T13:09:57
2026-03-19T13:10:57
...

One session per minute. For 12 hours straight.

Each session started with the same message:

[AUTONOMOUS TICK] You are running in continuous mode.
Check your goals, review shared memory for pending tasks,
and take any necessary actions.

And the response, 80% of the time:

NO_REPLY

The agent woke up, looked around, found nothing to do, and went back to sleep. But “sleep” still meant a full LLM round-trip: system prompt, tool definitions, the whole context window. Every 60 seconds.

The final breakdown:

• 169/170 sessions had the [AUTONOMOUS TICK] message
• 136/170 (80%) responded with NO_REPLY
• 35/170 (20%) actually did useful work
• One session ended with “You’ve hit your limit – resets 3am”

Finding the bug

OpenFang has a concept called “Hands” – bundled agent extensions with their own system prompts, tools, and settings. The Researcher Hand is one of them. Its config file sets max_iterations = 80, which is supposed to cap how many steps the agent can take in a single conversation.

But in kernel.rs, there’s this:

schedule: if def.agent.max_iterations.is_some() {
    ScheduleMode::Continuous {
        check_interval_secs: 60,
    }
} else {
    ScheduleMode::default() // Reactive
},

If max_iterations exists, the Hand gets put into Continuous mode with a 60-second tick. The code assumes “this agent can take multiple steps” means “this agent should run autonomously in a loop.” Those are completely different things.

8 out of 30 bundled Hands set max_iterations: browser, clip, collector, lead, predictor, researcher, trader, twitter. Activate any of them, and they start ticking every minute whether they have work or not.

Worse: activated Hands get persisted to ~/.openfang/hand_state.json. So every time the daemon restarts, they come right back. I activated Researcher Hand once through the dashboard, forgot about it, and it’s been ticking ever since.

What needs to change

The max_iterations field and the scheduling mode need to be decoupled. Something like an explicit autonomous = true in the Hand’s TOML file for Hands that actually want to run in a loop.

And idle ticks should back off. If the agent says NO_REPLY three times in a row, maybe wait 5 minutes instead of 60 seconds. Then 15. Then 30. There’s no reason to keep asking every minute when there’s obviously nothing happening.

I filed issue #756.

Takeaway

If you run OpenFang and have activated any Hands through the dashboard, check ~/.openfang/hand_state.json. If there’s anything in there you’re not actively using, remove it and restart the daemon. Otherwise those agents are waking up every minute, sending LLM calls into the void.

I set up an agent to research things for me. The biggest thing it researched was how much it costs to have an agent sit around doing nothing.