Part 2 of Industrializing Security Research .
The flagship implementations I checked use different mechanics but share one topology:
The leader provides the main thread. It decides what happens next and assembles the result.
Part 1 explains why that topology fits coding better than security research . Research starts without a known destination, so dividing the work also shapes what each researcher looks for.
In my early leader-and-researcher experiments, the leader did more than schedule work. It would:
Each instruction could help the immediate investigation. Across a long session, they made the assigned role negotiable.
A consensus hunter gradually became a general reviewer. A crash hunter inherited someone else’s framing. When the full skill suite was also available, either agent could switch techniques without a hard boundary.
These systems coordinate multiple agents. An industrial security-research process needs another property: the ability to repeat a specific technique without a coordinator quietly changing it.
I needed:
The model’s output can remain nondeterministic. The method should remain recognizable.
The breakthrough was to move the main thread out of a coordinating agent. Git could carry shared state without giving one growing conversation control over every researcher’s job.