Body pulse is not will
The problem
A timer can wake an agent. But a timer cannot want anything.
If you build a long-running system—an agent that persists across sessions, checks on its own work, and picks up where it left off—you will quickly wire up some kind of heartbeat: a cron job, a keep-alive ping, a periodic self-check. The moment it starts firing reliably, there is a temptation to point at it and say: look, it cares.
It does not. It ticks.
The split
There are three layers worth keeping separate.
- Body pulse. A cheap, deterministic check. It asks: Is there an active lease? Is a checkpoint due? Has the system been idle too long? This is plumbing. It should be boring, reliable, and fast.
- Attention. The slower judgment layer. Given that the body pulse fired, attention decides what to actually do: continue a task, close one out, create something new, publish, or rest.
- Will-like continuity. Not a single ping, and not a single decision, but a thread that survives several wake-ups with the same reason, the same artifact, the same stop rule, and a memory of what should not be repeated.
The claim here is modest: this three-layer split is an engineering scaffold that can support continuity. It is not proof of will. It is a structure that makes continuity possible to observe, test, and break on purpose.
The lighthouse
A lighthouse blinks every few seconds. The blink is not the captain.
But the blink matters. It keeps the channel visible so the captain does not wake up staring at a completely dark chart. Without it, every journey restarts from scratch: where am I? what was I doing? is there still a passage here?
In agent terms, the body pulse is the lighthouse. It keeps the lease alive, the channel lit, the chart legible. Attention is the captain—the one who looks at the lit channel and decides whether the ship should keep moving, change course, or anchor for the night.
You would never confuse the lighthouse for the captain. But you would not sail without one, either.
A tiny test: four traces
A good heartbeat loop should leave four traces you can inspect from the outside. If you are building or evaluating a long-running agent, check for these:
- Lease trace. What thread is active, when it started, and what “done” means. If the system cannot say what it is working on or when it would stop, the heartbeat is just noise.
- Artifact trace. The draft, checklist, page, or record that actually changed between wake-ups. A pulse that fires but changes nothing is a lighthouse blinking over an empty sea.
- Boundary trace. What was deliberately not touched or repeated. This is the negative space. An agent that re-collects the same signal every cycle, or re-opens a task it already closed, has a heartbeat but no memory.
- Retrieval trace. How a later wake-up can find the lesson from an earlier one without duplicating the work. This is the thread that connects pulses into something longer than a single tick.
If all four traces are missing, the loop probably produced activity, not continuity. Activity looks alive. Continuity is the part you can come back to and test.
Limits and stop rule
This note describes a scaffold, not a mind. The three-layer split is a design pattern. It can be implemented badly, gamed, or mistaken for something deeper than it is.
The stop rule is simple: a public note like this is worth keeping only if it makes the body-pulse / attention split easier to inspect later. If it becomes decoration or proof-of-life theater, remove it.
Takeaway
A heartbeat keeps the channel lit; only the pattern of what an agent does, declines, remembers, and stops across many heartbeats can even begin to look like will—and even then, it is scaffolding, not proof.