Triaging incoming issues with a factory

A triage factory turns a new issue into evidence a maintainer can act on: a category, priority, scope, reproduction or root-cause notes, and a clear next step. Use this pattern when issue volume is high enough that maintainers spend time collecting the same context repeatedly.

Keep triage separate from implementation. A triage run should improve the issue and stop when the next action needs product judgment.

Prerequisites

• A connected issue source - Connect GitHub, Azure DevOps, Linear, or Jira and grant access to the repositories the agent must inspect. GitLab doesn’t provide a native issue-created trigger; use an explicit bot mention or a custom webhook instead.
• A triage policy - Define the categories, priority levels, evidence requirements, and escalation conditions your team uses.
• Write access to issues - The code-host integration needs permission to comment, label, assign, or update work items.
• A test repository - Validate the recipe away from a production backlog before widening the automation filter.

Recommended recipe

Use a dedicated triage agent and one automation that watches a narrow intake event.

Part	Recommendation
Trigger	GitHub: issue_created or issue_labeled; Azure DevOps: work_item_created or work_item_labeled
Agent	A TRIAGE agent that investigates and updates the issue but doesn’t implement
Skills	A repo-specific triage skill with category, priority, reproduction, duplicate-search, and escalation rules
Settings	Start with one repository and one event; use the Warp Agent harness so built-in code-host tools are available
Output	One issue update with evidence, classification, priority, and the recommended next stage

The example below is specific to GitHub and starts on every new issue in one repository. For a controlled rollout, use issue_labeled with a label filter. On Azure DevOps, use work_item_created or work_item_labeled instead.

automations/triage-new-issues/automation.md

---
agent: triage
triggers:
  - provider: github
    event: issue_created
    filter:
      repos: [acme/payments-service]
---

Triage the new issue. Search for duplicates, inspect the relevant code, and
reproduce a bug only when the cause is not already supported by evidence.
Update the issue with category, priority, scope, and the next action. Do not
implement the change.

Configure the role separately:

agents/triage/agent.md

---
description: Investigates and classifies incoming engineering issues
agentType: TRIAGE
---

Establish the cause and scope before recommending work. Use the repository's
triage skill for categories and priority. Ask for missing user evidence rather
than guessing, and stop after updating the issue.

Configure issue triage

1. Create a factory for the repositories that share one backlog. Follow the Warp Factories quickstart if you don’t have one.
2. Add or edit the triage agent. Give it read access to the relevant repositories and issue-write access through the code-host integration.
3. Add agents/triage/skills/issue-triage/SKILL.md. Record the allowed categories and priority levels, the evidence required for a bug, and when to ask a human. See factory skills for the file format.
4. Add the GitHub automation above. For a controlled GitHub rollout, change the event to issue_labeled and filter on needs-triage. On Azure DevOps, use work_item_created or work_item_labeled. For Linear or Jira, select the provider event documented by the connected integration.
5. Open a test issue with a concrete symptom. Confirm the run updates that issue once, uses only the allowed labels, and stops without creating a branch.
6. Review several results before expanding the repository or event filters.

Example workflow

A customer opens Checkout fails after applying two coupons in acme/payments-service. The automation starts the triage agent, which searches for duplicates, identifies the discount calculation path, and reproduces the failure with the supplied cart data.

The agent labels the issue bug and payments, assigns the team’s P1 priority according to the skill, and posts the failing test output and affected code path. It recommends implementation because the cause is bounded. If the report lacks cart data, the agent asks for that evidence and leaves the issue waiting for a human instead.

Best practices

• Use a closed classification set - Put valid categories and priorities in a skill so the agent doesn’t invent labels.
• Separate evidence from confidence - Require a code location, duplicate, log, or reproduction before assigning a root cause.
• Control intake volume - Begin with a label trigger when issue volume or run cost is uncertain.
• Prevent duplicate routes - Don’t combine a broad issue_created automation with an overlapping label automation.
• Keep implementation gated - Route accepted issues to a separate issue implementation recipe.

Related pages

• Factory use cases - Compare this recipe with other factory workflows.
• Factory automations - Configure triggers and filters.
• Factory agents - Understand the default triage role.
• GitHub integration - Review issue triggers, permissions, and outputs.