Software, built by AI agents. Safely.
We move fast with AI agents. LumenFlow is the system we built to keep that work under control: it checks the tools we run through it before they act, saves their inputs and puts subprocess tools in a Linux sandbox where supported. We built it for our own work first.
Agents are brilliant. They'll also do anything you don't stop them doing.
Ask one to fix a bug and it may rewrite the payment system, read a password it was never given, or push straight to production. Telling it not to is a suggestion. We needed something that can't be talked out of.
We decide what a tool can touch before it runs.
For each tool call governed by LumenFlow, four permissions must agree: workspace, lane, task and tool. If one does not allow access, the kernel refuses the call. An agent cannot widen those permissions by asking.
Rules are rules, not suggestions.
For a governed tool call, one policy denial stops execution, however the agent phrases its request. Our delivery workflow adds branch, review and release checks around the code.
Inspect the work behind the result.
The kernel saves a governed tool's input before it runs, then tries to log the decision and result. A failed log write can leave a gap. Separately, LumenFlow Cloud links the receipts it receives. Signatures and outside timestamps are optional.
Subprocess tools run inside a real sandbox.
On supported Linux hosts, LumenFlow puts governed subprocess tools in an operating-system sandbox. They can see only the project files approved for that tool, and only declared environment variables pass in. In-process tools run in the host and rely on their own authority checks.
Several agents, side by side, all day.
Each agent has its own lane and copy of the code. Applicable checks cover formatting, types, tests and more before code work closes; broader checks run at release. A failed check sends the change back for repair.
A record of how the work moved.
Work-unit changes, handoffs and check results are recorded alongside the project. That history helps us see how often we ship, how long work takes, what breaks, how quickly we fix it and what comes back for another pass.
We use it to build it.
We use LumenFlow's scoped work, handoffs and applicable checks while building LumenFlow and LumenFlow Cloud. The kernel saves each governed tool's input before dispatch and attempts to log its result. Cloud links the evidence receipts it receives. These are products built inside the workflow we use ourselves.
Build the way we build.
Licensed · free and paid grantsPlayground open · sign-up closed for now