Factory Droid tutorial shows how to build AI dev teams

Fixing a minor typo or a localized bug is basically table stakes for modern coding agents. But handing an AI a massive database migration, paired with business logic refactoring and full test coverage, usually results in broken code and forgotten edge cases.

Factory is betting that the real battle in AI software development is no longer about raw chat completions, but the operating harness around the model. The company overhauled its Droid CLI to coordinate specialized subagents, manage parallel Git worktrees, and enforce rigid planning before touching production code.

Researching before writing a single line

Most coding agents fail on ambitious tasks for two reasons: they either start editing files prematurely, or they hallucinate vague implementation plans based on a shallow scan of the repo.

Factory tackles this with Spec Mode. Instead of jumping straight into the editor, Droid investigates the codebase and generates an engineering design spec outlining expected file changes, edge cases, and testing strategies. The developer reviews, challenges, and approves the blueprint while revisions are still practically free.

Once approved, these specs can be saved as Markdown files to serve as permanent documentation for future team members and other agents. The terminal interface also includes handy ergonomics, like hitting an exclamation mark on an empty prompt to drop into a direct Bash subshell, then pressing Escape to resume the agent session without polluting the context.

Parallel worktrees and scoped context

Real-world software lives in messy codebases filled with tribal knowledge. Factory attempts to tame this using structured context layers. Persistent rules live in AGENTS.md files placed at root or package levels, while task-specific workflows, such as deployment checklists, are packaged into dynamic Skills loaded only when needed.

Droid also supports Model Context Protocol servers to pull context from tools like Linear or Sentry, alongside headless execution via droid exec for CI scripts.

To prevent complex projects from stalling on single-threaded execution, Factory uses Missions to split work across custom specialist droids with different system prompts and tool permissions. Crucially, native Git worktree integration lets these agents run concurrently across separate workspaces, ensuring their file edits and test runs do not step on each other.

Why it matters

We think Factory's approach highlights where AI engineering is truly heading: frontier models are becoming commodities across Cursor, Claude Code, and Codex, meaning the workflow harness around the model decides practical utility.

For senior engineers and tech leads, Droid shines when the atomic unit of work is a full engineering initiative rather than a five-line patch. If you are executing dependency upgrades, API migrations, or multi-service syncs, the combination of enforced approval gates in Spec Mode and sandboxed worktrees provides guardrails that conversational chat interfaces completely lack.

The clever move here is designing for human oversight. By turning the agent's research into a reviewable design document, Factory respects how real engineering teams work. However, the barrier to entry is noticeable. Getting real value requires investing significant upfront effort into curating AGENTS.md, defining narrow MCP surfaces, and maintaining reliable test suites, which smaller, fast-moving teams might find too heavy for everyday hacking.