Refining agent
The most important agent in the system. Reads what came in — a one-line bug report, a rambling customer email, a stack trace — and turns it into a clean, classified, actionable brief, plus a confidence score that decides whether anything else happens automatically.
What it does
- Suggests a clearer title and description.
- Picks a type — bug, feature request, customer feedback, enquiry, task, or internal note.
- Sets a priority — low, medium, high, or critical.
- Suggests labels.
- Recommends an assignee from the project's members.
- Returns a confidence percentage on the rewrite.
- Spots "no action needed" submissions (compliments, thank-yous) and short-circuits them.
Settings
| Setting | Default | What it does |
|---|---|---|
| Model | The project default | Pick a different model just for this agent if you want to splurge or save. |
| Overwrite human edits | Off | By default the agent never replaces a title or description that a human has touched. Turn this on if you want it to always have the last word. |
How the workflow uses it
Every project starts with two rules built around the refining agent's confidence score:
- Above the confidence threshold, the brief is promoted to Ready automatically.
- Below the threshold, the brief is sent to a human for a second opinion.
You can change the threshold from Project → Workflow.
Confidence isn't difficulty. It's how sure the model is that its rewrite captures what the user meant. A short, clear bug report and a long, ambiguous customer rant can score the same — what matters is whether the refine is unambiguous.
What you'll see
On the brief detail page, the refining agent's work shows up as a "Compare / Diff / Reasoning" panel — original on one side, rewrite on the other, with the confidence number and a short note about what changed. You can accept, edit, or roll back any of it.