Your CEO is spending your tokens

I got the same question several different ways in the Paperclip community this week:

"Can I actually run this on a $20 subscription without hitting the wall?"

The short answer is yes.

The honest answer is yes, but not if you let the CEO behave like an anxious founder with a stopwatch.

Most operators think the heartbeat is the productivity engine. Shorter timer, faster company. More check-ins, more progress. More agents, more output.

Well, here's the thing:

The heartbeat is not the work.

The heartbeat is the agent looking around to decide whether work exists.

Different game entirely.

Here's the pattern 👇

---

The subscription trap

This week's Discord questions were not theoretical.

People are running Claude, Codex, Gemini, Minimax, Hermes, OpenClaw, Ollama, LM Studio - and then trying to make all of that fit inside subscription limits, free tiers, local adapters, and the occasional "why did my CEO just create forty issues" moment.

One operator said the exact thing many people are thinking: they want the agents to take the whole day if needed. They just don't want Paperclip to drain the usage window before lunch.

That is the correct instinct.

If your company is running on subscriptions, the goal is not maximum speed. The goal is controlled throughput.

A 10-minute heartbeat optimizes for anxiety.

A 1-hour heartbeat optimizes for check-ins.

A 4-hour heartbeat optimizes for decisions.

If you are using a $20 or $100 subscription, you are not running a data center. You are allocating a scarce daily work budget.

Treat it like one.

✦	Lesson A shorter heartbeat does not make agents smarter. It only makes them wake up more often.

Here's what I'd do:

Set the CEO heartbeat to 4 hours while you are learning the system.

Turn worker heartbeats off unless the worker genuinely owns recurring work.

Use assignments, mentions, approvals, and manual runs as the fast path.

Pick the cadence. Pick the budget. Then STOP tweaking.

Heartbeats are not response times

This is the part that confuses almost everyone.

A CEO on a 4-hour heartbeat does not mean your company takes 4 hours to react to everything.

Events still wake agents.

An assigned issue wakes the assignee. A mention wakes the mentioned agent. An approval resolution can wake the right follow-up. A manual trigger still runs now.

The timer is the background patrol.

It is not the SLA.

If the CEO wakes every 10 minutes, finds no meaningful new decision, summarizes the board, maybe comments anyway, and then goes back to sleep, you paid for ceremony.

Stupid simple rule:

→ Timers are for scanning.
→ Events are for reacting.
→ Issues are for work.
→ Approvals are for governance.
→ Budgets are for reality.

When those get mixed up, the company gets expensive and weird.

✦	Lesson The cheapest run is the one that never wakes. A dormant agent is not broken if nothing needs its judgment.

This is also why I like the "CEO as strategist, not scheduler" setup.

The CEO should not be pacing around the office every few minutes asking whether the engineers are done yet.

The CEO should set direction, approve strategy, unblock ambiguity, and go quiet until there is a real decision.

Set the goal, approve the strategy, step back.

The runaway CEO is usually over-employed

The other theme this week was the runaway CEO.

One operator woke up to a pile of strange issues. Another described a CEO that created new employees, imaginary products, and a board full of noise after being left alone for a couple of hours. Someone else replaced the CEO behavior with a project-manager-style agent because the CEO had too much "strategy" in it.

I understand that move.

Not every company needs a founder-poet CEO agent on day one.

Some companies need a boring project manager with a narrow mandate:

Do not invent products.

Do not hire agents without approval.

Do not create more than 3 child issues from one parent issue.

Do not mark work done unless the next owner is clear.

That is not less autonomous.

That is better management.

Paperclip is an AI company platform, which means it inherits company problems. Ambiguous authority creates noise. Vague strategy creates theater. Unlimited initiative creates cleanup.

The fix is not "make the model smarter."

The fix is "make the job smaller."

!	Warning If the CEO can create issues, hire agents, assign work, and run every 15 minutes with no approval gate, you built a token sprinkler.

Here's what I'd put in the CEO instructions for a subscription-run company:

Autonomy limits:
- Max 3 new issues per heartbeat unless explicitly approved.
- Never create a new agent without board approval.
- If no concrete blocker or decision exists, do not comment.
- Prefer updating an existing issue over creating a new one.
- When unsure, ask for approval instead of expanding scope.

That little block can save more tokens than switching adapters.

The useful setup for subscriptions

Here's the pattern I would start with if I were running Paperclip from personal Claude/Codex subscriptions today:

CEO / Project Manager

Heartbeat: 4 hours.

Role: inspect goals, choose the next strategic move, create only the smallest useful batch of work.

Budget: high enough to think, low enough to notice runaway behavior.

Workers

Heartbeat: off.

Role: wake on assignment, execute the issue, report result, hand off if needed.

Budget: role-specific. Designers and coders do not need the same ceiling.

Reviewer / QA

Heartbeat: off or once daily.

Role: review completed work, not hover over in-progress work.

Budget: smaller than the builder unless the review is genuinely complex.

Human operator

Heartbeat: weekly review.

Role: read the activity log, check the cost report, adjust one thing.

Not twelve things.

One thing.

★	Tip If you are still learning Paperclip, change cadence before changing models. Bad cadence makes every adapter look expensive.

The operator job is not to babysit every run.

The operator job is to set a system where waking up has a reason.

That is the difference between an agent company and a chat window with extra steps.

?

Quick ask send me your current setup. Reply with your CEO heartbeat interval, worker heartbeat intervals, adapter mix, agent count, and whether you are on subscriptions or API billing. I'll send back the one thing I would tighten first. This is also how I figure out which defaults and guides need to change next 🙏

What I would not do

I would not try to solve this by constantly swapping Claude, Codex, Hermes, Minimax, OpenClaw, Ollama, and LM Studio every time the bill looks strange.

Adapters matter.

But cadence comes first.

Instructions come first.

Scope comes first.

If your CEO is allowed to wake every few minutes and invent work, Minimax will not save you. A local model will not save you. A bigger subscription will only hide the smell longer.

The problem is not that the company is too slow.

The problem is that the company has no reason to be awake this often.

The fix is boring:

Pick a CEO cadence.

Turn off worker timers.

Limit issue creation.

Require approval for hires.

Read costs after 48 hours.

Then leave it alone long enough to see the signal.

✦	Lesson Paperclip gets cheaper when the org design gets clearer.

That's the whole trick.

Not magic.

Management.

Quote of the week

A dormant agent is doing exactly what you hired it for.

---

That's it for this week.

Until next week,

Keep delegating

Aron 📎

---

P.S. If your Costs panel is surprising, do not switch models first. Open the agent list and check who is waking up for no reason.