Use a native agent tool on its own, or add a workspace to organize projects and conversations. Choose the products yourself.
Foundation checks that the chosen tools can load the assigned skills, perform the work, support separate review and retain the results.
Updating an existing project? Use factory-foundation and its references/adoption.md: preserve local adaptations, select the reviewed revision, then verify loading and one real task.
The diagram shows the current Lab qualification: separate AgentOps, implementation and review VMs beside one another on Z13. It is a test layout, not a fixed number of agents or the final provisioning design. Trusted related projects may explicitly share an environment and its access.
Keep AgentOps pinned; workers are task threads. Our target is a separate sandbox per agent that AgentOps can create and manage from inside its own environment, with minimal physical-host actions and no extra coordinator LLM. Shared worker environments are an option with weaker separation. Automatic provisioning, cleanup, credential boundaries and GUI visibility must be qualified before adoption.
Run the complete workspace and agent tool inside the sandbox; keep real credentials outside it. Docker’s T3 Code guide provides a native SSH recipe. Choose scoped connections and verify them with Context & access.
Lab status — 28 October: AgentOps delegated a small task through implementation, two fresh reviews, repair and two new reviews in the prepared Z13 environments. Results returned natively to the same GUI. An approval needed operator intervention; this does not qualify unattended work. Two idle crash/restart checks passed. On-demand per-agent provisioning, approval handling and shared-host maintenance remain open in #32 and #34. Docker documents Ubuntu support; our Omarchy installation has no claimed vendor support, and custom secrets are experimental. The Mac is a client only.
AgentOps receives requests through the channels you choose. A personal Planning Agent can help shape a task before handing it over. Native events and schedules are optional.
Follow the owner's existing instructions about starting work and delivering it. Keep one lead per project; the Factory Lab develops the method.