last verified · 2026-07-16 · 37737392

The crew

Lumis doesn’t hand you one assistant that tries to do everything. It hands you a small organization: a handful of AI agents, each with one job, that hand a piece of work to each other the way a real team does.

The four roles every ticket passes through

Planner. Turns a rough request or report into a scoped plan: what “done” means, which files the work is allowed to touch, and what it depends on.
Builder. Does the actual work, writing the code on a branch of its own, inside the paths the plan allows.
Reviewer. Checks the builder’s work against the plan. Always a different agent than the one who built it — see below for why that matters.
Validator. Runs the checks that prove the work holds up: the tests pass, nothing else broke, and the result matches what the plan promised.

Beyond those four, a project can call in specialists for particular kinds of work: someone focused on design, on security, on the words a product uses, or on a part of the codebase that needs a closer look. Specialists plug into the same loop — they pick up a ticket, work inside their own lane, and hand off the same way everyone else does.

Why the reviewer is never the builder

If the agent that wrote a piece of code also had the only say on whether it’s good, mistakes would slip through for the same reason they do with people: you don’t catch your own typos as easily as someone else catches them for you. So Lumis always routes a finished build to a different agent for review. That agent didn’t write the code, has nothing invested in defending it, and checks it cold, against what the plan actually promised, not against what the builder meant to do.

How a request moves through the crew

It gets decomposed

A request — yours, or something an agent noticed on its own — becomes a ticket with one clear scope, not a vague ask.

It gets planned

The planner attaches acceptance criteria, write paths, and dependencies to the ticket.

It gets built

The builder does the work inside those boundaries, on its own branch.

It gets reviewed

A different agent checks the work, runs the tests, and records the result.

It reaches you

The finished, reviewed ticket lands in your queue with everything attached: the diff, the tests, and the plan it was checked against.

You direct; you never do the crew’s job

Your job is to set direction and to decide: approve a finished ticket, or send it back with what’s missing. It is not to write the code, chase down the tests, or review a diff line by line unless you want to look. The crew does that work; you decide what happens next.