Babysitter

The trait that decides it is that nothing is replaced. It wraps the coding agent your team already installed and already pays for, so adopting it changes how work proceeds without changing what anyone types, and abandoning it leaves your setup exactly as it was.

What it asks in return is that somebody writes the workflow, which is real design work and the part most teams underestimate. Pick it when the process matters more than the speed. Pick Plandex when you want the agent to decide the shape of the work.

The install matrix does not match the headline. A layer whose entire value is neutrality across coding agents has finished exactly two of them, and an experimental adapter in a workflow that enforces gates is worse than no adapter, because the enforcement is the thing you were trusting. Nothing states what experimental means in behavioural terms.

What it does right is stopping for people. Named breakpoints require a human answer before the run continues, which is a control most orchestration layers replace with a confirmation dialog.

1. This is the correct inversion. Rather than asking an agent to decide when it is finished, the process defines what must hold before progress is permitted, so the completion criterion is external to the thing being judged. 2. Expressing that in code rather than in prose makes it testable on its own.

2. What is absent is evidence that enforcement helps. No comparison of gated against ungated runs, no completion rates, no measurement of how often a gate catches something a reviewer would have missed.

The competitive position is squeezed from above. Every harness this supervises is adding hooks, gates and plan modes of its own, and when the two finished integrations ship equivalents natively the reason to run a wrapper thins out. There is no revenue to fund staying ahead of that.

Moat: none; the workflow definitions are portable by design, which is good for users and removes the switching cost. Likely path: the pattern is absorbed by the harnesses. Position: use it while the gap is real, and keep your workflow files simple enough to reimplement.

The journal is why this is on my list. When a regulator or a post-incident review asks what the agent did and who approved it, an append-only record answers in minutes rather than in a reconstruction from chat logs. That alone justifies the integration effort.

The cost is multiplied elsewhere: it drives whichever harness each developer already subscribes to, so my exposure sits on invoices I do not control, and it runs unattended, which widens that. Licensing itself is nothing. Approved with conditions: journals shipped to our own log store, and a spend cap on the underlying subscriptions.

Shipping both shapes is the right call. From my shell it drives whatever agent I point it at, and installed as a plugin it orchestrates from inside a session, so I choose where the control lives rather than accepting the author's preference. Permissive licence over readable code makes both paths mine to change.

What is missing is protocol support, so the servers I run are invisible to it and the workflow only reaches tools the harness already has. That is a wrapper's limitation more than a flaw.