Build a specialist once. Use it everywhere.
An agent is a named job with instructions, a model, the tools it may call and the shape of the answer it must return. Write it once, try it on real input, then reuse it in any workflow.
Support triage
v7 · released
Illustrative
- Instructions
- Read the ticket. Decide the queue and the urgency. Search the help centre before suggesting a reply. Never promise a refund.
- Model
- claude-sonnet-5
- Allowed tools
- knowledge.search linear.create_issue
From idea to specialist
Four steps, and each one is something you can check.
-
Name the job
Write one sentence about the result. “Route each support ticket to the right queue” is an agent; “help with support” is not.
-
Instruct and pick a model
Plain-language instructions, and a model chosen for the job rather than the biggest one available.
-
Allow only what it needs
Tick the tools it may call. Everything else stays out of reach, including other connections.
-
Define the answer and try it
Set the output schema, run it on real tickets, then release a version workflows can depend on.
Reuse
Improve it in one place. Every workflow that uses it can follow.
Workflows point at a released version of an agent. Ship v8 when it is ready and move each workflow across on purpose, instead of editing the same instructions in five places.
Inbound email
Help-centre chat
Weekly QA sample
Built in
What every agent carries with it.
Versions
Each release is numbered, and a workflow pins the one it was tested with.
Structured output
The answer is checked against its schema before anything acts on it.
Traces
Every call lands in Traces with its tokens and latency.
Scoped tools
It can call only the tools you allowed, through the workspace’s connections.
Start from a ready agent or describe a new role.
Test it once, then reuse it across the workflows that need it.