Ralph Orchestrator

The deciding trait is that it brings no model of its own. It drives the agent command you already installed and already pay for, so adopting it adds a supervisor rather than a subscription, and abandoning it leaves your existing setup untouched. That is a much easier decision than most orchestration tools ask for.

What you get in exchange is patience rather than intelligence: it keeps going, which helps on grindy work and hurts on ambiguous work. Pick it for long defined tasks. Pick Claude Squad when you want parallel attempts instead.

The failure mode is the design. Running an agent repeatedly against a plan until completion requires a completion test, and the documentation names none: no iteration ceiling, no external check, no cost bound. An agent that believes it is finished stops, and an agent that believes otherwise continues, on a backend that charges per attempt.

What it does right is observability of coordination. The specialised personas exchange events rather than sharing hidden state, so a stuck run can at least be read afterwards to see which one stopped making progress.

1. Repetition converts a reasoning problem into a sampling problem, which is a legitimate strategy and an expensive one: the expected spend depends on the success probability per attempt, a quantity nobody has measured here. 2. Verification is delegated entirely to the underlying agent, so the orchestrator inherits whatever checking that agent performs and adds none.

2. No evaluation is published. For a technique whose whole claim is that persistence beats a single pass, the absence of an attempt-count distribution is the missing number.

The strategic position is thin by construction. This adds supervision to command-line agents whose makers are all racing to add supervision themselves, and when one of them ships it natively the reason to run a separate process evaporates. There is no company, no price and no switching cost to defend the position with.

Moat: none. Likely path: the idea survives as a pattern and the implementation is superseded by a vendor feature. Position: use it while the gap exists, keep the plan files portable, and expect to stop needing it.

Nothing to license and nothing to negotiate, which sounds good until you notice what it multiplies. It executes without a person present, so it can run in a pipeline, and every one of those runs bills against a backend subscription somebody else on my team bought. Sixty developers times an unbounded loop is not a budget line, it is a surprise.

There is no administrative surface, no central log and no supplier. Onboarding is an hour for anyone comfortable in a terminal. Not yet, and not until spend is capped upstream of this tool.

The server mode is the detail I respect. One instance, one workspace root, over standard input and output rather than a port nobody closed, which means my other clients can drive it without opening an attack surface I then have to think about. Scoping is a taste question and this one has taste.

Rust binary, permissive licence, installable from the package manager I already use. Backends are swappable, so nothing here marries me to a vendor. I would keep this in my toolchain and I would read the source before I trusted a long run.