The principle: one function, not a new system
The main difference between a subagent and an "automation platform" is stated directly in the catalog: a subagent receives a limited fragment of context, executes one simple command and returns a structured result to a human or to the next stage. That means the implementation looks not like a migration to a new tool, but like the appearance of one new step in an already familiar process.
A working check of the principle: if after the implementation the employees had to change their whole working route — that is not embedding, that is rebuilding the process. Good embedding changes one step and leaves the rest untouched.
Five familiar routes
The catalog highlights five processes into which subagents are embedded without changing the working perimeter:
- Email. The Mail assistant sorts incoming messages and returns drafts to the email client or the work queue for review.
- Meetings. The Local meeting scribe turns a meeting recording into editable minutes with notes of uncertainly recognized places.
- Documents. The Legal RAG navigator answers about the uploaded contracts and hands the package of materials to the relevant specialist.
- Telegram. The Database consultant answers questions about metrics right in the messenger, in text and by voice.
- Requirements. The Requirements Analyst leads the materials to a specification and returns a document checkable against GOST and ISO.
Integration steps
The practical sequence is the same as for any of the nine subagents:
- Define the step. Which single step of the process eats time and has stable rules — sorting letters, minutes, contract search?
- Set the boundaries before launch. The sources, the allowed actions, the processing perimeter and the retention period are set before launch — this is an explicit requirement of the catalog.
- Agree the access. For example, the mail assistant is connected to the cloud, team or local perimeter only after access to the mailbox is agreed.
- Launch and check the result. A draft, minutes, an answer with links or a specification can be checked, changed and used at the next stage.
- Fix it in the passport. What the agent does, what it does not do, who confirms the result — eight fields of the passport.
What returns to the human
It is critical that the result is always a checkable artifact, not "the agent did everything": a draft of a letter, minutes with timecodes, an answer with quotes and references, a specification document. Every such artifact a human can open, correct and pass on. If a tool returns nothing checkable — it cannot be embedded into a process with human control.
Boundaries of embedding
Embedding does not cancel the limitations of the subagents themselves: letter classification can be wrong in context, the quality of minutes depends on the recording, a RAG answer does not constitute legal advice. The rule is one: the step into which a subagent is embedded is finished by a human — by checking, editing or confirming. More about how an agent differs from a subagent — in the article "Agent and subagent".
Questions and answers
Do I need to change my email client or messenger?
No: the result returns to the familiar route — the email client, the work queue or the messenger, and the sources and the perimeter are agreed before launch.
Can several subagents be embedded into one process?
Yes, if each has its own single function and its own access boundaries: for example, a meeting scribe at meetings and a mail assistant on incoming messages.
What to do if the process is not described yet?
First stabilize the rules of the step: a subagent receives limited context and one command, so an unstable process will give an unstable result.
Do the familiar tools break?
No: the subagent is embedded into the familiar process — email, meetings, documents, messenger or requirements stay in the usual route.
Where does the result return?
To a human or to the next stage of the process: a draft, minutes, an answer with links or a specification.
When are the integrations configured?
After the investigation: the sources, the allowed actions and the perimeter are set before launch.
What does the launch of a subagent include?
The passport, installation, basic configuration, work journal, and traceability start at $1,500. This price covers services; infrastructure and subscriptions are billed separately.
Can one subagent serve several teams?
The perimeter and the rights are configured per team: in the team perimeter access is separated by roles.
How to understand that the subagent took root?
By the journal and the accepted results: the step is stable, drafts go into work without a full rewrite.
What if the subagent makes mistakes?
An employee checks the result before the next step: mistakes do not go further, and the example returns for rework following the journal.