Zenith

The deciding trait is that it keeps looking for what is missing instead of announcing it is done. Anyone who has read a confident summary and then found three unimplemented cases knows that premature completion is the characteristic failure of long agent runs, and this is built specifically against it.

The cost of that is patience and tokens, and on a task you could have finished in ten minutes it is pure overhead. Pick it when the job is genuinely long. Pick something ordinary when it is not.

Self-directed expansion is the exposure. A session that concludes more checking is needed responds by starting more sessions, and the thing deciding is the thing being checked. There is no external referee in that loop, so a task the model finds unsatisfying can widen without a person choosing to widen it, and the invoice arrives afterwards.

What it does right is stopping discipline being treated as a named mechanism rather than an accident, which is more than the alternatives it was compared against manage.

1. The design of the study is defensible and the ownership of it is not independent. Five harness configurations were compared on a set of eight long-horizon tasks written for the purpose, with the baselines reimplemented by the same team, so both the difficulty distribution and the opposition were chosen by the party reporting the result. 2. Mean rank across eight items is a coarse statistic with wide error.

2. The isolated mechanisms are the useful output here, and they are stated clearly enough for somebody else to test.

Research output has a distinctive decay curve. It ships in excellent condition on the day the report lands, and then attention moves to the next question, because the incentive was the finding and not the software. There is no commercial entity behind this to fund maintenance and no customer to complain.

Moat: none, and none intended. Likely path: the mechanisms get absorbed into funded harnesses while this repository stops moving. Position: read the report, borrow the ideas, and do not expect a release next quarter.

Headless execution is the property that makes something reviewable as infrastructure rather than as a desktop toy, and it has that. What it does not have is anything my security team needs: no directory integration, no policy layer, no stated retention for what a multi-day run records along the way.

The platform gap rules out part of my organisation before we start, and a runtime that runs for days is a budgeting question I cannot answer at sixty engineers with the information published. Not yet.

Speaking two open protocols instead of one proprietary interface is what makes this worth my time. The orchestrator drives agents through their own channels, which means the thing being coordinated is my installation, with my keys and my configuration, and swapping one out does not require the project's permission.

One command starts it, the licence keeps a fork legal, and the whole control loop is Python I can read when it does something I did not expect. This is a harness in the honest sense of the word.