Always-On Is Not the Architecture

Persistent compute is the practice of keeping an agent's execution environment running continuously, whether or not work is occurring. It is the word doing the most work in the current wave of agent launches. xAI's Grok Bot gives every account a cloud computer and describes Bots that keep working 24/7. The framing is legible, and that is precisely its function: a teammate who is always at their desk is the easiest possible mental model to sell. But always at their desk is a product decision, not an architecture, and the two should not be confused.

The confusion is between two different kinds of persistence. What an agent needs to persist is its state: identity, permissions, memory, learned routines, relationships to tools and to other agents, and the exact position of unfinished work. What it does not need to persist is its compute. A business process spends most of its life waiting — for a reply, a document, a payment, a date — and an architecture that holds a running machine open through every waiting period is pricing attention rather than work.

The doctrine, restated

The Log encoded this distinction before the current launches made it topical. Event-Triggered Activation is the infrastructure pattern by which a dormant agent is initialised and executed in response to a defined external signal — a document arrival, a usage threshold breach, a time condition, a human action, or a downstream agent output — rather than running as a continuous process, making the compute cost of agentic operation proportional to actual execution rather than to elapsed time. Its complementary pattern, Suspend/Resume Architecture, allows an agentic workflow to pause execution at a defined checkpoint, persist its full execution state to a durable store, and resume from that exact checkpoint when a defined trigger fires. Together they close both boundaries — the agent not yet started and the workflow mid-flight — so that compute cost tracks work at every point in the process.

Read against that doctrine, the always-on products resolve cleanly. What they actually demonstrate is that persistent state has become a shipped default: the Bot's identity, credentials, files, and routines survive between sessions, and that survival is what makes delegation feel like delegation. The continuous machine underneath is a separate choice, and it is the choice the doctrine argues against. The likely convergence point is persistent state with ephemeral execution — the agent's identity never sleeps, its compute almost always does.

The economics of idle attention

The reason this is not a stylistic preference is cost structure. Infrastructure Is Architecture argued that infrastructure choices are business-model choices, not implementation details — and idle compute is where that argument bites hardest. A continuously running environment pays for compute during every idle period, and that cost accumulates in exact proportion to the time a business spends waiting, which is most of the time. At the scale of one prosumer account, a subscription absorbs it: the vendor charges a flat monthly rate, and the idle machine is a rounding error inside the margin. That is why the always-on framing is commercially rational for a lab selling seats. It is also why it does not transfer. An autonomous business running thousands of concurrent processes cannot flat-rate its own idle time to itself. Every hour of held compute that produces no execution erodes the Operational Arbitrage — the cost and output delta between a human-staffed operation and an equivalent agentic operation, widening over time as AI costs fall and human costs rise — that justifies the architecture in the first place. The arbitrage survives at scale only if waiting is free, and waiting is free only when pause is a first-class state.

There is a second cost, quieter than the first. A machine that is always on invites work to be assigned to presence rather than to signal. The Event That Wakes the Agent made the trigger specification itself the unit of design: defining precisely which event wakes which agent, with which context, under which threshold, is what makes an autonomous operation auditable and predictable, not an afterthought to be inferred from a machine that happens to be running. An always-on teammate that polls, watches, and putters is harder to reason about, harder to bound, and harder to bill to an outcome. The trigger specification is not bureaucracy. It is the operating contract.

The always-on launches are good news wrongly labelled. What shipped is durable identity, durable memory, and durable unfinished work — the state layer the doctrine requires. What was marketed is a machine that never sleeps — the compute layer the doctrine rejects. An operator adopting the substrate should take the first and decline the second: persist everything, run nothing until a defined signal fires, and let the cost of the business track the work of the business rather than the clock. Technology changes what is possible. Architecture determines what is affordable.