Writing
The Future of Engineering Resource Allocation: From Headcount to Compute Credits
Engineering organizations have always allocated ambition through headcount. Want more output? Get more people, more managers, more coordination overhead. Agentic engineering points at a different scarce resource: execution capacity, measured in compute, tool access, and agent runtime. I think of it as compute credits, and I suspect allocating them well becomes one of the more interesting organizational problems of the next decade.
Capital allocation, not headcount
The compute-credit model treats AI execution as an internal capital allocation problem. How much agentic capacity a person or team gets to spend — on exploration, maintenance, migrations, product bets — would depend on demonstrated judgment and past return, not just title. The promise is that individual contributors get to express leverage without first building a reporting empire. The unit of leverage shifts from people under management to resources under judgment.
But this only works if the organization can connect agent spend to outcomes. The unit cannot be tokens. It has to be work packets — cost, owner, expected value, risk level, measured result. Otherwise you reward the people who spend the most agent capacity rather than the people who create the most leverage, which is the same pathology as rewarding lines of code, in a new currency.
A mature version would look like an internal capital market. Teams get budgets, request more capacity for workflows with proven return, and report outcome data. Senior engineers earn more autonomy not because of tenure but because their track record shows they know which work is worth automating.
The controls you would need
Some boring machinery is required to keep this honest: per-workflow budgets with hard stop conditions, cost attribution down to the team and repository, outcome tracking so spend can be compared to shipped value, and risk weighting for anything that touches production, users, money, or policy. Boring is the point. Capital markets without accounting are casinos.
The failure mode to watch for is recreating headcount politics under a new name — credit hoarding, empire building denominated in GPU hours. Make the allocation rules explicit and the outcomes visible, or the old game continues with a new scoreboard. If AI really changes the cost of execution, it should change how engineering organizations distribute power. It would be a shame to get the first part and skip the second.