Revision: 0.1.10 · Updated: 2026-10-10
A practical method for building and improving applications with coding agents.
Use Factory when you want to give an agent an idea, a bug, a feature or an accepted backlog and get back a useful decision or a verified change. It gives owners and small teams a repeatable way to prepare a project, direct the work and review its delivery. T3 Code provides the workspace, native harnesses run the agents, and GitHub holds issues, pull requests and checks.
| Part | Responsibility | Example |
|---|---|---|
| Foundation | Prepare and maintain the conditions for work | Connect the execution host; establish context, access and real delivery checks |
| AgentOps | Direct work within the owner's mandate, before and during execution | Investigate an idea, select accepted work, coordinate workers and bring back a decision |
| ADLC | Perform and verify a concrete task | Specify a feature, implement it, independently review it and deliver under the agreed rules |
ADLC means Agentic Development Lifecycle. The skills work with your application's existing standards, tools and hosting choices. AgentOps coordinates ADLC and calls on Foundation when setup needs repair. These are responsibilities, not three servers or mandatory separate projects.
Keep one pinned AgentOps conversation in each application project. That lead owns the application's setup and work. The AgentOps conversation in the Factory repository develops Factory itself; it is not a central queue for every project. Share a relevant finding between projects when useful, with an explicit handoff.
| Scope | What must work | Existing guide |
|---|---|---|
| Execution host | The selected harnesses, accounts, access boundaries, skills and client connection; recovery and maintenance for the intended operation | Host setup |
| Application repository | Useful project context, reproducible setup, meaningful checks, standards and design continuity, GitHub rules, and the delivery/operation requirements of this application | Project context, engineering foundation, CI and delivery and GitHub preparation |
A working host does not make every repository ready. A well-prepared repository does not prove its agents have the right access. Reuse unchanged host evidence when adding a project, then check that project's actual tools and boundaries.
Foundation hands AgentOps a short result for the intended work: verified, required gap, or conditional/not needed, with the relevant source and proof. AgentOps starts authorized work on ready capabilities and returns setup gaps to Foundation. ADLC performs the accepted task and keeps affected project knowledge current. Reassess the affected capability when its code, configuration or access changes.
The shared baseline is useful context, explicit authority, working checks and observable access. Worker count, models, communication channels, hosting and extra isolation are installation choices. A plugin, prototype and production service need different preparation; none needs every optional control or a new collection of documents. Keep the result in the project's existing record.
Foundation also establishes a small configuration and adoption record. The project's AgentOps maintains its relevant lasting choices: model defaults and exceptions, skills, tool versions, update policy, connections and access. Link the actual configuration source, record why a change was authorized, and distinguish the intended setting from what was applied and verified. Keep sensitive host and account details private and credentials in their native secret stores. Changes are reviewable and reversible where supported; the record itself does not apply settings or grant access.
Factory's access architecture starts with the human owner and gives each agent only the access needed for its work:
- You, the owner, set the goals, approve or delegate decisions, and control which accounts, systems and permissions the agents may use. Personal and administrative access stays outside agent environments unless specifically granted.
- AgentOps directs the work within that mandate, assigns bounded tasks and brings back decisions. Any extra coordination or delivery access has a named purpose and scope; being the lead does not grant administrator rights.
- ADLC agents perform specific tasks in their assigned workspaces, using the tools and access needed to implement, test or independently review the change. They do not gain every permission held by the owner or lead simply by being delegated work.
Foundation prepares and checks these boundaries through the selected OS accounts, environments, native harness sandboxes and service permissions. It tests both an operation that should work and one that should be refused, and records the limits. Least privilege applies at all three levels; additional access can be configured for a concrete need and removed afterward.
Three access levels describe the intended responsibility and permission model. Claiming three isolated environments also requires technical separation. In a shared-account setup, AgentOps and workers can still reach the account's shared resources; separate threads, skills and worktrees do not change that. Use separately qualified identities/environments when workers must be unable to use the lead's resources. The architecture and setup guide explain the controls; the qualification record states what has actually been demonstrated.
You need a repository, an execution machine, access to the selected coding provider and permission to work with the repository on GitHub. T3 Code is the reference workspace; another capable harness can use the skills after its loading, access and execution behavior have been checked.
- Follow the setup guide to connect the chosen host and clients, authenticate the tools and adopt the skills from a reviewed Factory release. It also covers least privilege, updates and recovery. Its worked example uses a Linux host and a Mac client.
- Ask Foundation to assess and prepare your project. Reuse existing instructions, configuration and documentation; establish the missing context and real checks needed for the intended work. Record a small default model choice for the lead, implementation and review; AgentOps uses it unless you choose otherwise. The GitHub guide covers repository access, work tracking and delivery protections.
- Keep a pinned AgentOps thread in your project. Use the workspace guide to activate its role, then give it one bounded task and verify the result.
Give the lead an idea, a specific task or an accepted backlog. Examples:
- “Research this approach against our product goals. Recommend adopt, experiment, defer or reject. Bring me the decision before adding implementation work.”
- “Fix this accepted bug, run the relevant checks and get independent review. Bring me the candidate before merge.”
- “Work through these accepted issues in order. You may merge and release work that passes our rules. Ask me if new scope or significant risk changes the decision.”
There are two default human decisions: admit work to the implementation backlog, and accept its delivery before merge/release. The owner may delegate either decision for an explicit scope. Retain that delegation throughout the task. Tests and independent review still apply, and material new risk can require the owner to reconsider. See the operating policy.
A published method release is not an installed or qualified system. Verify your selected host, tools and a real task before relying on unattended operation. The qualification record separates package checks, behavioral exercises and live proof. Skills guide work; the native tools and infrastructure enforce access and run it.
See the architecture for responsibility boundaries and contributor guide for versioning, checks and releases. The earlier migration audit and research decision remain in release history; they are not prerequisites for a new installation.
Factory's method, code and documentation are MIT licensed. The Arc’ It AI name and logo, including their use in the banner, are not licensed for reuse by that software license. See brand asset terms. Existing application licenses, third-party tools and provider subscriptions retain their own terms.