When an agent writes the code, the prompt is the change.
PromptOps treats it that way: shared with the people who need it, routed, audited, and gated before anything ships.
Code has a pipeline. Prompts usually don’t.
Write, review, test, merge, deploy. Most teams send the input that drives all of it as a throwaway chat message.
| Practice | Unit of work | Focus |
|---|---|---|
| MLOps | Models | Training, serving and monitoring machine-learning models. |
| LLMOps | Model calls | Prompt engineering, evals and observability on top of LLMs. |
| PromptOps | The change itself | Prompts as team-owned, audited, reversible work that ships. |
PromptOps, in Orquesta.
Shared, with roles
Your team and your clients prompt the same agent. Each role decides what they can ask for and what needs review.
Routed
Each prompt is sized by complexity and cost, then sent to the right model tier — fast, balanced or premium.
Audited
Who asked, which model ran, what changed, when. Every run is tied to a person and a git commit.
Gated
The agent proposes, a human approves. Require sign-off on sensitive paths, or take control of the session live.
Anyone can ask. Your machine does the work.
- 01
Someone asks
A teammate or a client writes the prompt in the dashboard, Slack, Telegram or an embedded widget.
- 02
Your machine runs it
The agent on your server picks it up and works in the real repository.
- 03
Everyone watches
Output streams to the dashboard as it happens, logged against the prompt.
- 04
A person decides
Approve, send it back, or take the terminal and finish it yourself.
