Software Factory AI
This is the central concept. Everything else in the platform serves it.
ACE is a Software Factory AI. You describe software work in plain language, an AI runs the actual build-and-verify loop inside its own cloud sandbox, and a human approves the result before anything ships as a pull request. The AI is trusted to do the work: plan, build, and verify. It is not trusted to ship, merge, deploy, or spend without an explicit human "yes."
The idea
A traditional factory turns raw material into a finished product on a line, with people supervising rather than touching every part. ACE is that, for software.
- Raw material: your codebase, your cloud infrastructure, and a request in plain language.
- The line: an AI that plans the work, writes the code in its own sandbox, checks it, and prepares it for release.
- The supervisors: you and your team, standing on the loop, approving, correcting, and deciding at the points that matter.
The AI does the labor. Humans keep control.
"On the loop," not "in the loop"
This distinction is the whole philosophy.
- In the loop would mean a human is needed for every step: every line, every command. That is just assisted typing.
- On the loop means the AI runs the work on its own, and pauses to ask a human only at the decisions that carry real consequences.
ACE is on-the-loop by design. The AI is trusted to plan and build. It is not trusted to do anything hard to undo without an explicit human decision:
| The AI may do on its own | The AI must get a human "yes" for |
|---|---|
| Plan the work, break it into tasks | Finishing a feature (approval) |
| Write and change code in its sandbox | Merging a pull request |
| Run checks and health commands | Deploying, spending, deleting |
| Analyze a PR, an incident, a cost report | Anything it flags as costly or risky |
These gates are not reminders or suggestions. They are built into the platform. A feature cannot close, and a change cannot ship, until a person acts. See Feature Development for how the loop runs in practice.
The loop, in plain terms
- You ask. In plain language: a bug fix, a new feature, a question, from Chat or as a card on the board.
- The AI plans. ACE breaks the request into tasks.
- The AI builds. Specialist agents write the code inside a private cloud sandbox, never in your real repository yet.
- The AI verifies, and asks when unsure. It runs checks and, at each milestone or hard decision, stops and asks you for approval, for an answer, or for a go/no-go.
- You steer. You approve to continue, request changes to send it back for rework, or reject to stop it. You can also open the sandbox yourself and test the code by hand.
- You approve, and only then does a pull request appear. The finished work becomes a PR. The AI reviews the PR. You merge it.
Steps 3 to 5 repeat as many times as needed. That repetition is the inner loop: build a bit, check it, get human input, build more. The pull request is not the middle of the process. It is the end, produced once everything is approved.
Who does the building
You work with ACE as a single assistant, but behind that one identity the work is shared across specialist roles: a coordinator delegates, and specialists plan the work, build it, and review the result before it ships. If you want to see which role handled which part of a job, a card's Technical details view lays the run out for you.
How ACE splits the work is tunable per project, and you can shape how each role behaves. See AI Agents and Learn Path for how agents are shaped and grounded.
Why this matters
The Software Factory model is what lets ACE be genuinely useful and safe at the same time.
- Useful, because the AI does real, multi-step engineering work end to end, not just autocomplete.
- Safe, because every irreversible action routes through a human, and the AI works in an isolated sandbox until you approve shipping.
Two things ACE is not
- It is not a chatbot that only answers questions. It produces real code, in a real sandbox, that becomes a real pull request.
- It is not an autopilot. Nothing reaches your repository, your production branch, or your cloud account without a person approving it. ACE's cloud connections are read-only by design, as covered in Operations.