Beyond a prompt
An agent is more than a prompt pasted into a model console. Operationally, it is a versioned configuration: instructions, a model, an input contract, an output contract, optional tools, runtime limits, and a trigger. A control plane gives that configuration a durable home and links every execution back to the exact version that produced it.
That link matters when behavior changes. If a prompt edit overwrites history, an operator cannot explain why yesterday's run differs from today's. Runovio keeps one editable draft and treats a published version as immutable. Runs and schedules retain their version pin even after a newer version is published. Read more in the Agent Studio guide.
Control-plane responsibilities
A useful control plane separates configuration from execution. The UI and API validate and publish configuration; a runtime queues work, chooses the configured executor, applies retry policy, and records the outcome. Manual and scheduled triggers should enter the same path so they do not develop different validation or accounting behavior.
It should also make boundaries visible. A model cannot be allowed to choose arbitrary credentials, repositories, or URLs. In Runovio, GitHub tool IDs are allowlisted and a published version pins the installation and repository. Tool calls are serial and bounded. The model can request a permitted operation, but server-side policy decides what actually executes.
A practical boundary
The control plane should record enough to operate the system without turning logs into a second copy of sensitive prompts and outputs. Runovio keeps status, attempts, hierarchical spans, token counts, estimated cost, latency, and safe provider identifiers. It excludes complete prompts, provider bodies, raw GitHub evidence, and credentials from logs and traces.
This is deliberately narrower than a workflow marketplace or autonomous orchestration layer. The current product creates and runs individual agents, supports manual and scheduled triggers, and observes their execution. Start with the quick start, then explore agent observability and the runtime.