Pull the Cord

A lumberjack gets a chainsaw. Goes out into the woods and starts sawing — back and forth, like an axe. Slow. Awkward. His old axe was faster, he’s sure of it. This thing is terrible.

Two hours in, someone from the hardware store walks by. “Are you pulling the cord?” The lumberjack stares. There’s a cord. He pulls it. The chain starts moving. Oh.

This is what it looks like when you adopt a genuinely new tool and use it within the old paradigm. The lumberjack didn’t fail to use the chainsaw. He used it exactly as much as he could imagine using it — as a very expensive, awkward axe. The tool changed. The approach didn’t.

Engineers do this with agentic systems constantly. Teams get access to agent tooling and point it at the existing task list: features, bugs, code review. Nobody volunteers their agent capacity for dependency updates. Nobody schedules cleanup of stale branches. Nobody asks the agent to verify whether the documentation still reflects reality. That’s not negligence — it’s incentives. The fire is on the wall. You point the chainsaw at it.

The problem isn’t that agents don’t maintain things. It’s that nobody stepped back to ask whether the whole approach needed rethinking. Human engineering teams had an answer to this that wasn’t in any ticket: the senior engineer who’d been around long enough to notice the debt accumulating. The one who’d spend a Friday afternoon tracking down a flaky test or updating a dependency three versions behind — not because it was assigned, but because they’d seen what happens when it isn’t. That function existed outside the formal process. When agents replaced some of that function, the informal maintenance disappeared with it.

Capacity planning for maintenance needs to be explicit. Reserve a portion of total agent capacity — I use 15–20% — and allocate it to work with no immediate visible output: dependency updates, documentation freshness, test health, stale branch cleanup, session GC, hook verification. The work nobody volunteers for.

Not because it feels good. Because deferred maintenance compounds. Continuously maintained systems stay healthy at roughly linear cost. Neglected systems fail suddenly, and the repair is usually 3–5× what continuous maintenance would have cost. Every sprint you skip, the number gets worse.

There’s a second function to the reserved capacity: breathing room. The most expensive failure mode in agent systems isn’t “agent did something harmful.” It’s “agent did exactly what it was pointed at, and nobody checked the compass for three months.” The 20% that isn’t pointed at features is what lets you look up and ask whether you’ve been pulling the cord.

The 20% you spend maintaining the system is what keeps the other 80% working. Cut it, and the 80% starts shrinking.

Maintenance window for this article: booked in advance. Cord pulled before the first word.

Related reading:

• Most Rules Exist for a Reason — on why the maintenance rules are usually there for good reason
• I have No Memory Of This Place — on what happens when knowledge decays