Why Agents Do Nothing by Default

Why Agents Do Nothing by Default

Agents in Paperclip are dormant by default. They do not poll, they do not watch for events, and they do not run in the background consuming compute. Between heartbeats an agent is inert — no memory held in context, no budget consumed, no action taken. This is intentional design, not a limitation. Understanding why dormancy is the default is the first step to configuring agents that behave predictably and cost what you expect.

---

The problem with always-on agents

The intuitive model for AI automation is a process that runs continuously — always listening, always ready to act. That model works well for fast, narrow tools (a CI server, a webhook handler). It works badly for agents that reason, plan, and execute complex work.

Three failure modes appear quickly when agents run without bounds:

Cost spirals. A reasoning model left running continuously bills by the token. An agent that re-reads context, reconsiders its plan, and checks its task queue every few minutes accumulates thousands of dollars a month before producing meaningful output.

Noise. Always-on agents tend to over-produce: too many status updates, too many partial attempts, too many messages in task comment threads. Operators stop reading them, which defeats the audit trail entirely.

Unpredictability. If an agent can act at any moment, interventions become racing conditions. You edit an instruction file while the agent is mid-run. You update a task priority while the agent has already checked it out. Predictable behaviour requires predictable execution windows.

Dormancy solves all three problems at once. An agent that runs only when triggered is cheap between runs, quiet when there is nothing to do, and easy to reason about because each execution window is a discrete, logged event.