The score hangs on the scaffold

Whenever an agent runs, there is machinery outside the model. Which tools it is shown and in what order. How finely a single action is cut. What comes back when a step fails, and how it is asked to recover. The authors call this external machinery the harness — scaffold is the closer everyday word.

The awkward part is that building the scaffold up does raise the score, and the gain stays tied to the scaffold. Worse, the best scaffold differs across domains, across instances and across models. Anyone building a general-purpose agent has to pick a side: settle for one shared scaffold that is optimal nowhere, or keep adding specialized ones and route each case to the right one. The first discards performance; the second accumulates machinery that never stops growing.

The core: use the scaffold as the teacher, not the runtime

The proposal reduces to one move. Take the optimized scaffold and use it as training-time guidance rather than deployment-time machinery. Transfer the behaviours it induces into the weights, and remove it at deployment. The authors call this harness distillation.

The obstacle sits right there. The optimized scaffold and the target scaffold differ in action space and in what information is available. Guidance produced under one cannot serve directly as supervision for the other, because the vocabulary of actions is not the same. What bridges that gap is agent-as-harness: having an agent, rather than code, play the role of the harness. Guided by the optimized scaffold, that harnessing agent corrects the student's responses before execution, and the corrections are expressed in the target scaffold's action space. Guidance thereby turns into training demonstrations. Fine-tuning on the resulting trajectories is meant to internalize behaviour the scaffold used to draw out.

What the abstract states

The authors report experiments spanning knowledge work, tool use and science. Three claims are placed in the abstract.

That is the extent of the abstract. These are the authors' claims, not settled conclusions. The second one in particular runs in the direction of "better without the scaffold than with it," which warrants care. As a preprint, the comparison conditions and the treatment of counter-evidence wait on the full text and on review.

What it looks like translated into practice

Shipping the scaffold along

Procedure documents, the ordering of check items, the template for sending work back. A full set built for each domain, distributed to each site. Results come, but every new domain means another set. Someone new is needed to decide which set applies to which case. Every set needs its own maintenance.

Shipping without the scaffold

The built-up set is used only while training. What goes to the site is the trained state itself. Only one thing circulates. But what was internalized is not visible from outside. The grounds for acting a given way can no longer be pointed at in a document.

The difference is not speed or cost. It is whether the procedure sits outside or has dissolved inside. While it is outside, it can be read, corrected and pointed to in an audit. The moment it dissolves inward, all three get hard. Distillation trades explainability for portability.

What is not new, and what cannot be read

Transferring behaviour from a stronger side to a weaker one is not new. Distillation has a long history, and turning tool-use behaviour into training data is by now unremarkable. What this paper contributes is placing the target of the transfer on the behaviour the scaffold induced rather than on the model's outputs.

A great deal is unreadable from the abstract. How strong a model plays the harnessing agent. What share of responses received corrections. What level the target scaffold was fixed at. None of these conditions is given. The result that removing the scaffold beats keeping it also leaves an alternative reading open: that the attached scaffold simply did not suit the setting it was measured in. Nor is it stated at what granularity the recovered behaviour patterns were counted.

Votes and implementations arrived; a cluster did not

The paper drew 35 upvotes after posting, and the public repository drew 28 stars. A number of readers judged it worth reading, and a number of implementers went to run it. Meanwhile, no other paper on the same question appeared in the same window within the range collected.

Two signals stand — readers and implementers — and the cluster of papers does not. This site's selection measures the spread of a subject across independent signals: votes, implementations, coverage and clustering. Neither the votes nor the stars is a measure of correctness or importance. They say some people thought it worth reading and some thought it worth running, and nothing more should be loaded onto them.

Where this connects to pharmaceutical and regulatory practice

What the paper touches is an old question in regulated work wearing new clothes. Does the procedure sit outside the person, or inside them? Structurally this is the same question training practice has always faced.

In material review, outside means the check items are written down so that anyone can walk them in the same order. Inside means an experienced reviewer notices something is off without consulting the document. Practice needs both. But only the outside version can be produced as a record. "It felt wrong from experience" is not usable language in an audit.

So bringing this approach across requires splitting the question in two. Removing the scaffold for speed and portability is fine on its own terms. But it must not be removed from any step whose grounds have to be pointed at. Separate out in advance which judgements carry accountability, and dissolve inward only what lies outside that boundary. Postpone the separation and what emerges is a high-performing step that cannot state its reasons. In regulated work, the harder object to handle is not the mechanism that gets things wrong — it is the mechanism that gets things right and cannot say why.