A firm of agents that has reinvented itself over the past fifty years: it anticipates disruptive changes but builds on the strengths it already possesses

Out of 24 periods assessed, the judge consistently ranks the company’s clarity above its investment decisions — an average difference of 1.9 points out of 10. And the score rises as objectivity falls.

Multi-agent systems based on language models are increasingly organized like businesses — assigned roles, internal criticism, shared memory — but are evaluated on tasks that last a few minutes. Bravish Ghosh reverses the time scale with Frontier Autolab: a simulated company, incarnated by sixteen role people and a dedicated red team, has to refound in nine technological eras from 1990 to 2040. Each time is temporally divided: the council decides from a dated note, a historian judge then reveals what really happened and notes the decision on five dimensions, and the lessons are put into a persistent notebook.

On four trajectories — 36 decisions of the period, 180 sub-notes — a gap returns without exception: in the 24 periods evaluated against real history, the judge notes the recognition of the coming shift over the choice of the layer to invest, with an average difference of 1.9 points on a scale of 10. The reason is prosaic: the boards choose the layer that their existing assets allow them to reach. The organizational design is also sustainable. A red team with digital shutdown thresholds produces fifty years of drivers under conditions and no products — and it is this company that is the first class rating grid. Firms whose memory contained lessons in market structure rotate at each time; the one whose memory retained only validation procedures retains a single method from beginning to end.

The most interesting is the doubt that the author returns against his own test bench. The score increases over time in each execution, while the historical regression sub-note attributed by the judge decreases — correlation of -0.58 within the same trajectory. The apparent learning is therefore confused with the memorization of history, and other distortions are attributed to self-assessment, selection of briefing notes and aggregation of scores. Archive and API harnesses are published, with the specification of the fictitious and post-cut periods that would make this device a real reference bench.