Automating the software development lifecycle with a factory

An end-to-end software development lifecycle (SDLC) factory carries a work item from intake through triage, planning, implementation, review, and verification. Use it after the individual stages work reliably on their own and your team has explicit human gates.

For repositories that ship together, use one factory with several agents. Splitting one product by stage creates extra routing and ownership problems.

Prerequisites

• Reliable stage recipes - Validate issue triage, issue implementation, code review, and any required Computer Use verification before connecting them.
• One source of truth - Choose the issue tracker and code host that hold status, acceptance criteria, pull requests, and review history.
• Protected human gates - Decide who approves specifications, accepts incident risk, and merges pull requests.
• Repository-specific skills - Record test commands, review policy, release checks, and escalation rules.
• Scoped credentials - Give each role only the repositories, secrets, and MCP servers it needs.

Recommended recipe

Use the default role agents with the foreman as the entry point. Let the foreman choose the shortest safe path rather than forcing every request through every stage.

Stage	Agent	Expected output
Intake and routing	Foreman	Work-item ownership, questions, and the next stage
Triage	Triage	Evidence, scope, priority, and open questions
Planning	Spec	Product and technical plan with validation criteria
Implementation	Implement	Code, tests, validation, and a draft pull request
Review	Review	Independent findings and a recommendation
UI verification	Verify, when needed	Observed behavior and visual evidence
Merge	Human	Approval and merge under repository policy

Route approved issues to the foreman:

automations/approved-work/automation.md

---
agent: foreman
triggers:
  - provider: github
    event: issue_labeled
    filter:
      repos: [acme/payments-service]
      labels: [factory-ready]
---

Own the approved issue through human handoff. Triage only when cause or scope
is unknown. Require a human to approve specifications for material behavior
changes. Implement and validate the change, request independent review, add UI
verification when the acceptance criteria are visual, and never merge.

Add follow-up automations for review feedback and pull request completion only after the main path is stable. The default GitHub connection creates automations for labeled work, submitted reviews, and closed or merged pull requests.

Configure the SDLC workflow

1. Create one factory for the repositories that ship together. Enable the foreman, triage, spec, implement, and review agents.
2. Configure each role with the narrowest required access. Use a separate verify agent on the Warp Agent harness when the product has a rendered interface.
3. Add shared repository conventions under skills/, then add role-specific procedures under agents/<name>/skills/.
4. Connect the issue tracker and code host. Choose one ready-work signal, such as factory-ready or a Ready for Engineering state.
5. Add the intake automation above. Test clear small work, ambiguous product work, review revisions, and a failed validation path.
6. Confirm the foreman pauses for specification approval and questions, while repository protection prevents agent merges.
7. Add metrics and Scorers only after the workflow produces consistent outputs.

Example workflow

A GitHub issue labeled factory-ready reports that invoices omit a purchase-order number. The foreman sends the issue to triage because the affected service is unclear. Triage identifies the rendering path and confirms the scope.

The change affects an external API contract, so the spec agent writes validation criteria and waits for a maintainer to approve them. The implement agent updates the API and tests, then opens a draft pull request. Review finds a missing migration check and sends the work back. After revision, the verify agent exercises the invoice preview and attaches evidence. The foreman hands the reviewed pull request to a maintainer, who approves and merges it.

Multiple-factory boundary

The factory definition format doesn’t provide a cross-factory dependency graph or an automatic merge stage. Use one factory for one product’s SDLC. Use several factories when products and repositories have separate ownership and release cycles, with each factory running its own lifecycle.

If an external process must hand work between factories, use explicit code-host or tracker events, custom webhooks, or Factory MCP. Those handoffs aren’t atomic: define idempotency, failure ownership, and the system that records overall state. Keep automated merging outside the recipe unless your repository policy and an external orchestrator explicitly own it.

Best practices

• Earn each stage - Add orchestration only after the individual stage has a measurable output and a clear owner.
• Skip unnecessary work - A small, well-defined fix doesn’t need a specification.
• Keep status in durable systems - Use the issue and pull request as the record, not an agent transcript alone.
• Design failure paths - Decide what happens when validation fails, a person doesn’t answer, or an integration is unavailable.
• Measure outcomes - Track accepted pull requests, revision causes, cycle time, and repeated failures before optimizing models or prompts.

Related pages

• Factory use cases - Start with one stage before adopting the full workflow.
• How Warp Factories work - Review the work-item lifecycle and human checkpoints.
• Factory agents - Configure roles, models, harnesses, and access.
• Measure and improve - Add Scorers, benchmarks, and controlled self-improvement.