PDM

AI agents ​

PDM can be connected to an AI assistant — Claude Code, say, or any other application able to work with external tools. Once connected, the assistant gets access to PDM data and can answer questions about stands, settings, releases and builds on its own, and run operations when needed.

We call such a connected application an agent. It does not act "on its own": the agent talks to PDM in your name and strictly within your permissions.

Why bother ​

Answering a question like "where is the new version of the service already deployed?" normally means opening several stands one after another and comparing values by eye. An agent does that for you: it walks the data itself and answers in a single sentence.

An agent can reach stands and their availability, configuration and service versions, platform applications and their actions, releases and builds — exactly what the PDM interface shows you. Nothing beyond that is opened up to it.

Example requests ​

Requests are phrased in plain words — there is no special syntax:

  • "Which stands have version 1.42 of the api3 service deployed?"
  • "Show me how the settings of stands prj-demo1 and prj-demo2 differ."
  • "Which stands are down right now, and for how long?"
  • "What went into the latest release and how many stands has it been rolled out to?"
  • "Find every stand where this parameter has an override."
  • "Build the feature/login branch of the ui2 service."

The first five are read-only: the agent changes nothing. The last one already leads to a real action in the infrastructure, and requests like that follow separate rules (see below).

Limits: what an agent can and cannot reach ​

An agent never gets more permissions than you have. It acts on behalf of the user who approved the access. If you cannot see a stand in the interface, neither can the agent. If you are not allowed to change the configuration of a production stand, the agent will not change it either, however nicely you ask.

Which leads to something non-obvious: an empty answer from an agent does not always mean "there is no data". Sometimes it means "no permission". If an answer looks suspiciously empty, check whether you see the same thing in the PDM interface.

A change to your permissions changes what the agent can do straight away — there is no need to reconnect it. See Permissions for how permissions work.

Data reflects the latest synchronization

PDM does not own the infrastructure, it reflects it: information about stands, configuration and secrets is read from external systems on a schedule. An agent's answer shows the state as of the last synchronization, not necessarily as of this very moment.

Operations that change something ​

An agent can update a stand's configuration, run an application action or a build, create a release.

Such operations follow a separate rule: before making a change, the application shows exactly what it is about to do and waits for your explicit approval. Without confirmation the operation does not run. Read-only requests need no confirmation.

Changes reach real environments

A configuration update, a release, an application action or a build means commits in a repository and pipelines that run against live stands, production included. Before confirming, check which stand is actually meant: stand names repeat across clusters, and "the same" stand easily turns out to be from the cluster next door.

Be especially careful with secrets: if you are allowed to read them, the agent can obtain decrypted environment variable values. Ask for them only when you genuinely need them, and remember that from that point on those values live in your conversation with the AI assistant.

Long-running operations and "the command was sent" ​

Long operations in PDM run in the background: the request returns immediately while the work proceeds step by step. The interface receives progress by itself, but an agent gets no live event stream — to learn the outcome it has to ask for the operation state again.

The practical consequence: when an agent says "the command was sent", that is exactly what it means, not "the deploy has finished". Ask it to check the state a while later, or look at the workspace page yourself.

Responsibility and revoking access ​

Everything an agent does is recorded against your account. Therefore:

  • do not connect PDM to an application you do not trust;
  • read what is being proposed before you confirm it;
  • revoke an agent's access once you no longer need it.

A connected application occupies a separate session tagged “Agent”; ending it deprives the agent of access immediately — see Active sessions.

What's next ​