Plan mode: the assistant shows the plan before it changes anything
How the assistant built into Opfield handles real changes: it researches with read-only tools, shows a plan, waits for your confirmation, runs it in the background and checks the result.
By the Opfield team4 min readUpdated
Opfield has an assistant built into the interface, AI Workspace. You tell it what you want to reach in plain words, and it works out the steps. For work that changes infrastructure, it first shows you a plan and waits. Nothing changes until you confirm. Below is how that works and where the limits are.
Describe the outcome
You can start in two ways: write the outcome you want, or pick a guided Scenario. Scenarios cover common tasks: setup, infrastructure health, nodes, Docker, publishing and diagnosing routes, TLS, logging, notifications, databases, status pages and delegating access. A free-form request works the same way. "Release payments-api 2.4.2 with no downtime" is enough to begin.
For ordinary questions the assistant simply answers. It explains how something works, reads the documentation or looks up the state of a resource. It knows which page you are on, so "why is this route failing" means the route in front of you.
Research before any change
When the work has several steps, touches several resources, needs a lot of research or carries real risk, the assistant switches to Plan Mode. You can also turn Plan Mode on yourself.
In Plan Mode the assistant has only read-only tools. It looks at the current state: the release, logs, metrics, routes and the nodes involved. If something is unclear, it asks a clarifying question instead of guessing. Answering a question does not authorize anything else.
The new version expects STRIPE_KEY, which the production environment does not have.
A plan you confirm
The research ends in a structured plan. It keeps apart what was found, the review of intent and safety, the steps to carry out and a final verification.
Actions that change anything become available only after you confirm the plan explicitly. Until then the assistant cannot change a thing, however the conversation goes.
Review a plan the way you would review a colleague's change. It should name the exact resources, not "all affected services", and say what customers will notice, when to roll back and how the result will be checked. If a step is a shell command or a direct database edit, ask whether a regular Opfield operation can do the same.
Running in the background
A confirmed plan runs in the background. You see progress step by step and can pause, resume or cancel it. The run belongs to Opfield, not to your browser tab, so closing the panel does not stop it.
When the steps are done, a separate verification run checks the result. A tool call that succeeded does not mean the goal was reached, so the plan ends with that check, not with the last step. It is still worth testing the customer-facing path yourself afterwards.
Each Work Session has one active plan at a time. Separate Work Sessions can run their plans independently.
Your permissions, nothing more
The assistant has no access of its own. It acts for the user who asked and works only through Opfield tools that check permissions. The model gets only the tools your scopes allow, and Opfield checks scopes, what your license includes, validation and resource state again for every operation. If you cannot do something in the interface, the assistant cannot do it either.
Planning is separate from approval mode. Planning is always read-only. A confirmed plan then follows your current approval settings, so actions that need your approval still ask for it. The assistant also respects per-user tool access. It asks before it reveals or rotates credentials or exports private keys, and it cannot change your sign-in methods, MFA or sessions. Scenarios follow the same rules.
Every action in the audit log
The assistant uses the same operations as the interface, the REST API and MCP, and its actions land in the same audit log under the user who asked. To see what a plan changed, open the audit log and the tasks it created rather than relying on the chat summary.
If you prefer to work from your own agent in an editor or chat app, the same tools and rules are available over MCP, as described in Running infrastructure with AI agents through MCP.
Optional, with your choice of model
AI Workspace is off by default. An administrator turns it on and chooses the model: a configured OpenAI-compatible provider or a published Opfield Inference model. No data goes to an AI provider before that. The model is pinned to each conversation, and the assistant warns you before you switch models mid-chat.
Everything the assistant does is also available without AI, in the Operations Console, the REST API and MCP. Chat is available on every plan. Plan Mode and Scenarios need Personal or higher.
Getting started
Ask an administrator to turn on AI Workspace and connect a model, then start with a read-only question about a service you know well. The AI Workspace page at https://docs.opfield.dev covers permissions, data handling and a checklist for reviewing a plan.