GILD verdict
Standardize on Factory's Droid if your engineers need models from several vendors behind one agent, under central policy in every place they work.
Put model policy, an autonomy ceiling, and review discipline in place before usage spreads. Then test Factory's vendor-reported benchmark results on your own repositories with the models you would pay for.
Factory
AI Coding Agent by Factory
+22 PositiveFit and limits
Who it fits
- You want models from several vendors behind one agent. The catalog spans Anthropic, OpenAI, Google, and xAI, plus Droid Core open-weight models, at per-model rates, and Factory Router can pick the model per task.
- Engineers work in different places: the Droid CLI, the desktop app, VS Code, Cursor, Windsurf, JetBrains IDEs, or Zed, and you want the same agent and the same managed settings in each.
- Your security team needs the agent loop to run on your own machines. Droid reads and edits files locally without uploading or indexing a copy of the repository, and deployments range from Factory Managed to a Custom Deployment without a Factory-hosted control plane.
- You have large, separable projects such as migrations. Missions plan features and milestones, run workers with validation, and can continue for hours on Factory-managed Droid Computers.
- You want a harness that adds accuracy on a model you choose. On the Terminal-Bench 2.0 leaderboard, which Stanford University and Laude Institute run, Droid scored 77.3% with GPT-5.3-Codex. Terminus 2, the benchmark's own reference agent, scored 64.7% on that model.
Where it falls short
- Individual and Teams usage runs on three rolling rate limits at once. Each model draws at its own rate, from 0.04× for GPT-6 Luna to 12× for GPT-5.5 Pro, with Claude Opus 5.5 at 1.6×. Factory sizes Plus at about 5 times and Max at about 10 times Pro usage, and publishes no token allowance for any plan.
- Software Factory and Droid Shield 2.0 are in Private Preview, and Droid Shield 2.0 is enabled by request to Factory support. Delegations from Slack, Linear, Jira, and Microsoft Teams work on any paid plan, and a Slack connector lets Droid search channels and messages inside a session.
- Source control integration covers GitHub and GitLab. Self-managed GitHub Enterprise Server or GitLab requires Enterprise, and the documentation lists no Bitbucket or Azure DevOps connection.
- Factory's headline benchmark results are its own runs, and Droid's newest Terminal-Bench entry dates from February 24, 2026. Hands-on reviews by practitioners range from strong to failing.
Implementation reality
Set model and key policy first
Enterprise Controls decide which models and gateways are allowed and whether users may bring their own keys. Org-level hard controls win, and lower levels can extend a policy but never weaken it. With direct enterprise endpoints, Factory routes model traffic straight to the provider. With your own gateway, the gateway's privacy terms apply.
Cap autonomy and sandbox commands
Autonomy runs Off, Low, Medium, or High, and an org-wide maximum caps every user and project. Permission rules allow, ask about, or block shell commands, and the OS-level sandbox limits file writes to the working directory and network access to allowed domains. Factory's own guidance keeps High autonomy inside containers or VMs.
Put a ceiling on usage
On Business and Enterprise, usage limits come from the contract, and on Enterprise, User Groups apply a monthly limit to every member of an identity provider group. The Analytics API and dashboards report consumption by model and by user, and Factory Router routes routine steps to cheaper models.
Choose the deployment pattern
Cloud-managed runs Droid on laptops and CI with Factory cloud as the control plane. Hybrid keeps model and telemetry traffic inside your network. Fully air-gapped needs no runtime connection to Factory, and Factory Private, announced September 18, 2026, puts the control plane in your VPC or data center. EU organizations keep session content and inference in Europe.
Keep a human on every merge
Shared delegations open pull requests as the Factory Droid GitHub App, so the requester can approve a pull request Droid authored. Factory's documentation recommends a branch rule or status check that requires a second human approval on those pull requests.
What to measure in your pilot
- Run Droid and your current agent on the same 20 closed tickets with the same model; standardize on Droid only if its merged changes beat the incumbent's.
- Have each pilot engineer run /limits at the end of every day for two weeks. Move a seat to Plus or Max if it hits the 5-hour limit more than twice a week.
- Log who approves each pull request a shared delegation opens for the first month; if requesters approve their own, require a second human reviewer before widening access.
Sources
- Official documentation
- 23
- Public source
- 1
- Factory: Pricing
- Factory docs: Individual Plans
- Factory docs: Organization Plans
- Factory docs: Available Models
- Factory docs: Droid CLI
- Factory docs: Factory App
- Factory docs: IDE Integrations
- Factory docs: Delegations
- Factory docs: Software Factory
- Factory docs: Factory Missions
- Factory docs: Droid Computers
- Factory docs: Factory Router
- Factory docs: Autonomy Level
- Factory docs: Sandbox
- Factory docs: Enterprise Controls & Managed Settings
- Factory docs: Deployment Patterns
- Factory docs: Data Flows & Privacy
- Factory docs: Self-Managed Source Control
- Factory docs: User Groups
- Factory news: Introducing Flexible Deployment Options (Sep 18, 2026)
- Factory news: Droid, the #1 Software Development Agent on Terminal-Bench (Sep 25, 2025)
- Factory docs: Droid Shield
- Snorkel AI: Terminal-Bench 2.0 leaderboard, Stanford University and Laude Institute (as of Oct 25, 2026)
- Factory docs: Feature Maturity