Pimp My IDE / Garage Dispatch
Back to garage
September 21, 2026 | agent runtime / lifecycle / cost

Idle is a state, not a free tier.

Agent infrastructure now wants explicit verbs for active, suspended, resumed, and destroyed work. That is useful. It also creates a new contract: what survives the park, what wakes the worker, and when parked state gets scrapped.

THE TAKE: Do not bolt suspend and resume onto an agent as cheap parking. Define the snapshot boundary, wake authority, lease, and destruction deadline before you count the idle time as savings.
Work the Idle State Tow Rig

The old job model has a clean finish.

Kubernetes describes a Job as a one-off task that runs to completion and stops. The controller can retry failed Pods, track completions, and clean up the Pods when the Job is deleted. Suspending a Job deletes its active Pods until the Job is resumed.[1]

That contract works when completion is the natural boundary. An interactive coding agent is messier. It may wait for a model, a tool, a person, or a future event while its filesystem and process state still matter.

AX gives the waiting state a verb.

Google's new AX project declares Tasks, Workspaces, Gateways, and Models through versioned manifests. Its command line adds suspend and resume beside the familiar apply, get, watch, and delete operations. The project says a suspended task can continue with its state intact.[2]

The repository also carries a blunt warning. AX uses a v1alpha1 API, its core ideas are still changing, and breaking changes are likely before a stable release. Treat the lifecycle model as a useful design signal, not a production recommendation.

A parked agent still has cargo, keys, and a return ticket.

The runtime is betting on idle time.

AX runs on Agent Substrate. That project maps many stateful actors onto a smaller worker pool because agent-like programs spend much of their time idle. Its README reports sub-500 millisecond resume operations, more than 500 suspend or resume activations per second, and a demo that multiplexes about 250 actors over eight Pods.[3]

Those numbers are project claims from its own benchmark and demo. Agent Substrate also says it is in early development, is not ready for production, and offers no compatibility guarantee. The stronger lesson does not need the benchmark. State-preserving parking changes the unit you must operate.

Write the parking contract.

  1. Snapshot: list the memory, filesystem, process, credential, and connection state that survives. Name what is rebuilt instead.
  2. Wake: name the event or person allowed to resume the task. Add deduplication and a retry limit before one signal becomes a traffic jam.
  3. Lease: cap parked age, storage, spend, and retained authority. A suspended task should not keep an open-ended claim on secrets or money.
  4. Scrap: define expiry, deletion proof, and the handoff path for work that must be recreated.

Now suspend has an operational meaning. It is neither stopped nor free. It is a leased state with a known wake path and a known end.

Interactive makeover / lifecycle contract

Idle State Tow Rig.

Traditional purpose replaced: a binary running or stopped badge. Better version: move the task through active, parked, resumed, and scrapped states while the snapshot, wake, lease, and destruction contract stays visible.

Tow the task through a named lifecycle

This teaching rig does not inspect a runtime or prove that any state was saved or deleted. Its readout comes only from your selections.

Lifecycle state
Parking contract
CONTRACT OPEN0 / 4 INTERLOCKS
Parking contract selections

Active does not mean bounded.

The task is shown as active, but no lifecycle evidence has been selected.

Why it is better: state, retained cargo, wake authority, lease, and disposal share one receipt. Even with every interlock selected, the status stays "review gate ready" because the rig never talks to the runtime.
Sources read, not vibes
  1. Kubernetes documentation, "Jobs": one-off completion, retry behavior, cleanup, and the effect of suspending a Job.
  2. Google AX repository and AX project page: the Task, Workspace, Gateway, and Model resources; suspend and resume commands; network fencing; stated scale goals; and the explicit pre-stable warning.
  3. Agent Substrate repository: actor-to-worker multiplexing, lifecycle operations, benchmark claims, demo scope, architecture notes, and the warning that it is not production-ready.
  4. Hacker News discussion, item 49780797: the exact discovery thread for AX. It is a reaction source, not proof of performance or readiness.

Source boundary: AX and Agent Substrate are connected projects. Their benchmark and scale numbers are first-party claims. Kubernetes documents a different workload contract and does not validate those claims. The tow rig is a planning aid, not production telemetry, a cost estimate, or proof that suspension is safe.