dots Coordinating Claude Code Through FlowRMM

Getting useful work from multiple AI tools creates a coordination problem. Someone must carry the goal between them, inspect changes, follow tests, and decide the next step. FlowDevs is testing how much coordination we can give to OpenAI's dots while keeping work visible and permissions bounded. In our lab experiment, dots worked with Claude Code on a Windows Runner machine through FlowRMM. dots inspected the desktop, assigned work, checked the response, and tracked the change through testing. Claude Code handled the coding. FlowRMM provided the governed path to the computer. The goal is a useful engineering loop that can eventually make approved progress overnight. We have completed individual cycles, and the longer-running reliability test remains underway.

How dots Coordinates Engineering Tasks

A coding task moves through several states: investigating, editing, waiting for tests, responding to a failure, preparing a review, and checking what shipped. Each transition needs context. In this workflow, dots kept the larger assignment in view while Claude Code worked inside its development environment.

It inspected a fresh screenshot, determined whether Claude was still working or needed direction, and sent the next scoped instruction through FlowRMM. The loop is inspect, assign, implement, test, review, and verify. The result of one cycle becomes the input to the next.

A sent instruction is one checkpoint. A passing test or a verified release is another. Keeping those checkpoints separate prevents a successful click from being mistaken for completed engineering work.

The Permission Boundary Stays Explicit

A human approved a 12-hour GUI-control lease for one Windows console in this test. That lease gave the workflow a defined computer, control surface, and time window. The scope covered GUI input. It did not provide blanket shell authority or permission to deploy to production. Work outside that scope still needed its own authorization.

FlowRMM connected observation and approved action. dots could inspect screenshots and dispatch governed desktop actions, then check the resulting state before continuing. Human priority and the ability to pause remained part of the workflow.

Authentication has a separate job. A working OAuth connection establishes access to the service under its permissions. It does not replace the human approval of a particular control scope. These boundaries matter more as a session gets longer. The workflow needs to preserve them through waiting, reconnecting, and resuming.

What the Workflow Has Actually Completed

One completed cycle addressed an OAuth issue. Claude Code implemented the fix, and 105 focused OAuth tests passed. The change also went through CI and database checks, the normal merge queue, and a successful portal-only lab release. dots followed the work through those stages and checked the outcome. After verifying the release, we restored the pause on automatic lab deployments.

That is a specific engineering result in a lab environment. It gives us evidence that this coordination path can carry a real change through implementation, checks, and a bounded release.

A separate Claude Code documentation session, using its own worktree, produced four documentation changes and passed 53 focused tests. A pre-existing Flutter documentation-checker issue was also reported. Those changes are now in a draft pull request. Independent review found no new blocker; the pull request is unmerged and CI is still pending.

Both outcomes belong in the record. The tests that passed tell us something useful. The outstanding checker issue and pending CI tell us where work remains.

Building Toward Overnight Iteration

Overnight coding needs continuity as well as useful code. A long-running session has to survive the ordinary interruptions of real software work: an expiring connection, a slow test, a changed screen, or a task that needs a decision. It also has to recognize when its approval no longer covers the next step.

We are actively testing connection renewal and post-expiry behavior. A completed overnight run has not yet been demonstrated in this experiment. The next proof is whether the workflow can continue across those boundaries, retain the right context, and leave a clear record of what it completed, what failed, and where it stopped for human input.

Success would mean waking up to reviewable changes and trustworthy results, with unresolved decisions clearly identified. We still have to prove that longer-running behavior.

Frequently Asked Questions

Does dots have full production access?

No. The control grant covers GUI input on a specific Windows console. It does not provide blanket shell authority or permission to deploy to production.

Has FlowDevs proven autonomous overnight coding?

Not yet. We have completed individual cycles in a lab environment. We are actively testing post-expiry renewal and overnight reliability, but a completed overnight run has not yet been demonstrated.

What role does FlowRMM play in this system?

FlowRMM provides the governed computer-control path. It connects observation with approved desktop actions while maintaining explicit permission boundaries.

Why Coordination Matters for Real Work

This experiment shows the kind of operational software we build at FlowDevs: connected systems in which people set direction, specialized tools do their work, and the handoffs remain visible.

dots supplies coordination. Claude Code supplies implementation. FlowRMM supplies the governed computer-control path. The human supplies the goal and the authority to proceed. The practical value is reducing the manual work of carrying context from one tool to another while preserving the evidence needed to review the result.

We have a working loop and concrete lab results. Now we are testing how reliably that loop can continue over time.

Explore FlowRMM at https://flowrmm.com. To see how FlowDevs connects software, people, devices, and AI into working operational systems, visit https://flowdevs.io or schedule a conversation at https://bookings.flowdevs.io.