Engineering on an Exponential, Drawn

Engineers used to plan for ten-year timelines. Now agent capability roughly doubles every few months, and it is tempting to shrink every plan to the next release, because that is as far as anyone can see. That is backwards. A senior engineer's value is seeing further than the next release, and on an exponential the long view is harder and worth more.

The way through is to separate two things. Capabilities (which model, what it can do, which scaffolding) change in months: engineer them on short, cheap, reversible cycles. Invariants (data durability, security boundaries, cost and energy, trust and accountability, regulation, how organizations change, interfaces) hold for a decade: engineer them on long horizons, because they constrain every future agent however capable it becomes. Step through each diagram with the buttons or arrow keys; every number is dated and linked to its source, with the caveats.