Execution Flow

At a high level, execution moves through these stages:

1. an operator or API client launches a workflow
2. the platform creates activation and task state
3. a runtime claims work under an explicit contract
4. the runtime prepares workspace, tools, credentials, and execution environment
5. the agent loop runs
6. results, artifacts, and logs are reported back to the platform

This is the basic loop to keep in your head while reading the rest of the docs. Agirunner is not “the model doing work directly.” It is a control plane creating work, an execution plane claiming it, and a shared evidence layer carrying the result back.

What The Platform Does

• creates the work
• chooses the role and execution contract
• exposes logs, artifacts, and operator actions

The platform is also the place where later activations can be triggered. If a workflow changes because of a result, a failure, an approval, or operator steering, the platform decides whether the orchestrator should wake up again and what it should see.

What The Runtime Does

• prepares execution
• runs the loop
• executes tools
• captures output
• reports status back

That means the runtime owns the mechanics of doing work, but not the meaning of the work. It can report “this task completed with these artifacts and these logs.” It does not redefine the workflow contract that led to the task.

What Operators Experience

An operator mostly sees this flow through the dashboard:

• launch from a playbook
• watch the workflow board update
• inspect live console and deliverables
• intervene if the workflow needs attention

The underlying architecture matters because those operator surfaces are only useful when the system has clean records of launch, activation, task execution, and result capture.