The Window

Predictions should be falsifiable. Here are mine, dated.

Year	What happens
2026	The first funds run agent workforces in production, quietly, the way the early quants were quiet. Underway.
2027	"Describe your AI research process" appears in allocator due-diligence questionnaires. The early adopters stop being quiet, because the results stop letting them.
2028	Coverage depth becomes marketing. Two-person funds publish research at institutional depth. The headcount question in allocator meetings inverts from "how many analysts" to "how does your workforce run."
2030	The pyramid visibly inverts. The best-known research operations in the industry are small senior teams running large workforces, and the entry-level analyst role as we knew it is a case study.

If 2028 arrives and allocators are not asking, I was wrong, and this page will still be here.

The order of operations

For a desk that moves now, the sequence matters more than the speed.

One: write your process down. Agents run processes, not intentions. The desk that cannot articulate what it actually does, step by step, name by name, will automate nothing but its confusion. This is the unglamorous part, and it is also the moat, per Essay V. The act of writing the process down is itself worth the price of the exercise.

Two: start with monitoring. The highest return and the lowest blast radius. Watching what you already own carries no thesis risk, no sizing risk, and no career risk, and it is the fastest way to learn what a workforce feels like before you trust it with anything louder. Nobody's first deployment should be discovery. Everybody's first deployment should be the position they were already supposed to be watching.

Three: own the workflow, rent the models. Model providers commoditize each other with enthusiasm, and the desk should let them. The workflow, the thing tuned to your book, your process, your definitions of what matters, is the asset. Never let it live on someone else's side of the wall.

Four: move before the questionnaire asks. Once the question is on the DDQ, the answer is table stakes, and table stakes are not an edge. The return on this transition is highest for the desks that complete it while it is still optional.

The asymmetry

Price the two errors honestly. The cost of being early is small, capped, and measured mostly in time. The cost of being late is structural, compounding, and eventually visible to your LPs in the only font they read.

Every repricing in this industry's history rewarded the early and taxed the late, and every single one looked optional from inside the window. The terminal was optional. The internet was optional. Alternative data was optional. The window is open now. Windows in this business have a habit of becoming walls.