Skip to content

Canvas

Open it: Canvas · under Agents in the console menu.

A canvas says where and how one flow runs once it is deployed: in a shared pool or a sandbox of its own, on which runtime, how many copies, whether it remembers between runs, and what happens when something it needs is down. Until it is deployed, an agent or flow runs on the shared runtime inside FlowR, and needs no canvas.

One canvas deploys one flow. An agent’s canvas deploys the agent: the agent’s own flow, with the agent’s settings.

A preset is a canvas with no flow: a ready-made way to run. To deploy, you copy a preset and bind the copy to your agent or flow.

PresetHow it runsWho may use it
isolated-ephemeralIts own sandbox, a fresh one per run; nothing carried between runsAdmin in the project
isolated-persistentIts own sandbox; each run checkpoints, and the next resumes from itAdmin in the project
session-persistentFor agents that must remember between triggers: state is checkpointed and resumes on the next callAnyone who may deploy
shared-ephemeralIn the shared runtime; no state carried between triggersAnyone who may deploy
prod-sharedProduction, shared pool: 2 to 5 copies, no dedicated runtimeAdmin
prod-haProduction: a dedicated runtime, 2 to 10 copies, scaled on queue depthAdmin

“Admin in the project” includes a project’s owners and developers; a person with an Operator account who owns the project is Admin there.

The quickest way: on the agent, open Versions & deploy, pick a preset under how should it run?, and click Deploy. A3T copies the preset into the agent’s project as the agent’s own canvas, binds it to the agent, and deploys it. The agent then shows deployed on its runtime, and its own runtime pod takes its work.

Undeploy on the same tab stops it and removes its pod; the agent goes back to the shared runtime. Whoever may change the agent’s project may deploy and undeploy it.

Open a canvas from Items. Six tabs:

Canvas Details: name, project, description, the flow it deploys, status

FieldWhat it means
Name, Project, DescriptionWhat it is and where it belongs
FlowThe one flow it deploys. Empty: it is a preset
StatusWhether it is deployed, and where

Canvas Placement: shared pool or isolated sandbox, runtime labels, pinned runtime, and where it lands

FieldWhat it meansEmpty means
Where it loadsShared pool: beside other work. Isolated sandbox: its own pod, nothing else beside it; for untrusted code, noisy work, or data that must stay apart—
Runtime labelskey=value pairs, comma-separated, that a runtime must carry to take it, e.g. class=gpuThe default runtime
NamespaceThe Kubernetes namespace its pods go inThe runtime’s own
Pinned runtimeOne runtime by name. It must still match the labelsPicked by labels

Where it lands lists every runtime and whether it takes this canvas, so you see the result before deploying.

Canvas Lifecycle: state and session, scaling, restarts

SectionFields
State and sessionRunning or Paused (hibernate). Ephemeral: nothing carried between runs. Persistent: each run checkpoints and the next resumes from it, for agents that remember
ScalingAt least and At most copies; Scale on a metric such as queue_depth (empty: CPU); Target per instance
RestartsRestart on failure, always or never; Backoff exponential, linear or fixed; At most so many restarts

Canvas Dependencies: what feeds it, what it calls, resources it depends on, and what happens when one is down

SectionWhat it shows or sets
What feeds itThe triggers that send to its flow
What its flow callsOutside systems, read from the flow’s definition
Resources it depends onResources watched while it runs. Empty: FlowR watches the resources its flow calls
When one is downqueue: hold new runs until it is back. cordon: take no new runs here; runs elsewhere go on. circuit-break: fail fast for a while, then try again. fail: fail the runs that need it

The canvas as a graph (where it loads, what it depends on) and as the YAML Studio edits.

Canvas Deploy: what it will do, and how it is running

Deploy shows what will happen (the flow, the runtime and sandbox, how it runs) and lists anything to fix first, such as a preset with no flow. Live shows how it is running once deployed.

Runtimes shows each deployment on its runtime. From there, or from the agent’s Access › Controls, you can pause it (callers get the paused response), resume, restart or undeploy it.