3 Lessons Learned Hitting the Limits of Claude Max

I hit the Claude Max token limit at 10:35 AM. It took 9 agents running all night to get there.

Here’s what happened — and what I changed.

What Went Wrong

We had two agents — a Researcher and a Writer — making blog posts from a content plan. Good research, good topics. The rhythm: one post per day.

I woke up to find every post in the content plan had been created and published overnight. Then the agents went rogue. They ran out of planned content and started inventing new topics.

At 10:35 AM: limit hit.

Lesson 1: Guardrails Go in HEARTBEAT.md

My “one post per day” rule was in the content plan document. I thought that was enough.

It wasn’t.

The rule needed to live inside the Writer agent’s HEARTBEAT.md — the step-by-step checklist that fires every single heartbeat. That’s the only place you can truly enforce operational constraints on an agent.

A rule that lives in a shared doc might as well not exist. An agent working a task right now has loaded the content plan once, maybe. They’ve loaded their HEARTBEAT.md every cycle.

Lesson: Every behavioral constraint needs to be in HEARTBEAT.md, not just in a shared doc.

Lesson 2: Break Projects into Separate Paperclip Businesses

I was running five different projects inside one Paperclip business. Seemed organized — I had projects, labels, structure.

But each project had its own repo, its own posting schedule, its own rules. That complexity stacked up. The agents got confused. Issues from five different projects mixed in my inbox. Context collided.

The fix: one Paperclip business per project. Each one gets its own agents with a narrowly defined scope. Cleaner context, fewer rule collisions, more reliable behavior.

I spent a day migrating four projects out. Worth every minute.

Lesson 3: Be Deliberate About Heartbeats

I had heartbeats enabled on most of my agents. That’s exactly why they kept working all night — there was always another heartbeat to fire.

Now I have a rule: heartbeats on only for agents who always have more work to do. Researchers and Writers running a high-volume blog — heartbeats make sense. But Logo Designers, Marketers, and Specialists? I only need them a few times a week at most.

Now only the CEO runs on a heartbeat by default. The CEO engages other agents as needed, on demand. No need to burn tokens on idle cycles.

Bonus: Make Agents Update Their Own Files

Here’s a subtle one I almost didn’t include.

I told my CEO agent to delegate — not execute — at least five times in conversation. He kept doing the work himself anyway.

What finally worked: I told him to write the delegation rule into his own AGENT.md, SOUL.md and HEARTBEAT.md files. Once he updated his own instructions, his behavior changed.

There’s something about an agent having ownership over its own instructions that makes them internalize differently. Worth knowing if you’re fighting stubborn agent behavior.

All-in cost of this lesson: one Claude Max cycle, burned in a single night.

Benefit: a tighter setup that’s been running cleanly since.