Subagents

A subagent is a separate helper inside your AI for a narrow job: a code reviewer, a debugger, a technical writer. Each has its own prompt, its own role and — most importantly — its own set of tools, the things it is allowed to do. onplate keeps subagents next to skills and installs them into the AI tools that support them with one button.

Tools and permissions

The power of a subagent — and its risk — is in the tools it was given. onplate shows the set right on the card and scores it:

Read-only (Read, Grep, Glob) — the subagent can look, but not change and not run. Safe.
Write (Write, Edit) — it can change files. Deserves attention.
Execute (Bash) — it can run commands. The highest risk.

An empty tool list means the subagent inherits all the permissions of the AI — also a reason to look closer. onplate flags such cases and puts them into the “capabilities” part of the security check.

The security check

The body of a subagent is a set of instructions the AI will follow, so it goes through the same scanner as skills: prompt injection, secret leakage, dangerous commands, hidden unicode. On top of that, the tool set is scored. The result is an overall score (Safe / Suspicious / Dangerous), a numeric risk and a confidence, and every finding carries its evidence line and a “what to do” recommendation. More in the Security section.

The Security scan button checks every subagent at once; adding a new dangerous subagent asks for a confirmation.

How to add one

Create subagent — set the name, the role, the model and the tool set by hand;
Import — paste the content, or point at a file or a folder;
From repository — pull subagents from a connected Git repository;
from the Marketplace — on the Subagents tab, with a preview and a security check.

Installing into an AI

On the card pick the scope (global or project) and the target AI, then press Install — or All at once. Subagents are supported by the AI tools that have a subagent mechanism (Claude, GLM, opencode). As with skills, the exact path is shown before writing, and every install is visible in Bindings.

An install creates the subagent profile in the config of the chosen AI, but it does not start the subagent. Starting happens later, from the AI tool itself.

How to check the result

After a single or a bulk install, Orbi shows which profiles were confirmed for each target and which could not be installed. On deletion it also distinguishes deleting from storage only from removing the profiles from disk.

If Orbi reports that the final state is unconfirmed, the operation may have applied partially. Refresh Subagents and the Subagents tab in Bindings, check the targets and paths, and only then repeat the action.

If you delete a subagent from storage only, the profiles already written keep working and become unmanaged. Restoring the record from the trash does not recreate the profiles that were removed from disk earlier.