Prepare the repository once, plan each change in Spec Mode before any edit, set an autonomy level that matches where the session runs, and review the result against tests before anything merges.
When you need this
- Engineers have started using Droid and the quality of what comes back varies by person.
- Sessions run long and wander, or edits land before anyone agreed on the approach.
- You want one team standard for planning, autonomy, and review.
Evaluation criteria
Before work: repository instructions
Where do build commands, test commands, and conventions live?
Look forWrite Droid's commands and repository rules in AGENTS.md at the root. Missions workers inherit the file, so one file governs a session and a multi-agent run.
Before work: readiness
Can an agent verify its own work in this repository?
Look forRun /readiness-report. Factory's Agent Readiness Model has five levels. A repository reaches a level after passing 80% of the previous level's criteria. Factory recommends Level 4 for Missions. Use /readiness-fix to address missing checks, such as linting or unit tests, before delegating larger work.
During work: Spec Mode
Did anyone approve the plan before code changed?
Look forUse Spec Mode for changes larger than a small fix. It plans read-only and waits for approval before implementation. One Hacker News builder has one model write the spec and another implement it, so that no model can rewrite code to satisfy its own bad specification.
During work: autonomy level
What can Droid run without asking?
Look forSet Low for edits on developer laptops. Use Medium for package installs and commits. Reserve High for containers or VMs, as Factory advises. Permission rules can still block a command or require approval at any level.
During work: session size
Is this a session, a custom droid, or a Mission?
Look forUse a normal session until the task needs parallel work. Create a custom droid for a recurring specialist job. Use a Mission when work splits into milestones with a clear integration step. Factory estimates at least one run per feature plus two validator runs per milestone.
Before merge: review
Who checks the change, and against what?
Look forRun tests and CI on every change. Use /review for a first local pass, then assign a human reviewer. Droid Shield blocks commits and pushes through Droid when the staged diff contains likely secrets. It does not scan Git commands run outside Droid.
Questions to ask vendors
- Which repositories have an AGENTS.md and a readiness level we have checked?
- Which changes skipped Spec Mode last month, and how did they review?
- What autonomy level does each environment allow, and who set it?
- Which repeated corrections should become an AGENTS.md rule or a skill?
Red flags
- High autonomy on a laptop with standing production credentials.
- Missions started on repositories with no scriptable way to run and test the app.
- Large diffs merged on a green build with no named human reviewer.
- Secrets pasted into prompts when they should be read from environment variables.
Products in this piece
Sources
- Official documentation
- 9
- Public source
- 2
- Factory docs: AGENTS.md
- Factory docs: Agent Readiness
- Factory docs: Autonomy Level
- Factory docs: Permission rules
- Factory docs: Factory Missions
- Factory docs: Missions Planning & Validation
- Factory docs: Droid Shield
- Factory docs: Security
- Factory docs: Deployment Patterns
- Sid Bharath: Factory Droid Tutorial: Missions, Skills, and Agents (Sep 30, 2025)
- sebazzz on Hacker News: one agent writes the spec, another implements it (Feb 16, 2026)