SWE budgets: from Opex to Capex · Jay HackAll writing

The increasing autonomy of code agents has fundamentally changed what is possible in software engineering. Large scale operations that were previously invalid "chess moves" are now a matter of budgeting: the Bun-to-Rust refactor is a watershed moment, demonstrating a total tech stack overhaul for just over $150k in compute. This will soon become commonplace (to Anthropic's delight) and it's a beautiful thing - will create better software overall. With it, the way we think of SWE budgets will evolve towards capex.

Until now the discussion has been dominated by the tokens-per-head model, with individuals achieving higher personal limits with seniority. I think we are imminently headed towards a scenario where SWE compute is treated like a fixed/shared resource, however, that is apportioned across competing projects, as opposed to individuals. The role of developers will shift: establishing what is important to build in the first place, "owning the outcome" and playing the role of ultimate verifiers - although each of these steps will also draw heavily from centralized compute budgets. The top SWEs will move to companies that allot them greater moral sway over compute allocation.

There's a near parallel to this in ML research, at larger firms like Google - compute hours are allocated via an internal marketplace dynamic. But while this apparently has created a toxic, political snake pit inside of Google, SWE is different in that outcomes are fundamentally verifiable, payouts are more clear, and therefore the decision process is more straightforward/less political. A Rust refactor for a large API surface, for example, can be modeled ahead of time and the payoff more clearly knowable. The "politics" will unfold in determining which metric to improve, which is more aligned with hierarchical priority setting and less susceptible to political influence.

As compute prices fall and the high water mark of autonomy/capability rises, the frontier of what is possible will grow with it. The marginal contribution of a human sitting in front of a synchronous chat session will drop and the overwhelming leverage will reside in making smart decisions about where to direct (growing) budgets for swarms of agents.

Playing this out a few years - with the acknowledgement that very few SWE-related things I do today cannot be achieved via some sort of MCP - I think we arrive at the level 5 self-driving CTO in short order. Technical outcomes will be a much more direct function of capital expenditure and traditional bottlenecks around communication/culture will fade to the background. There's an astronomical amount of value to be created and captured in this transition and even more once any organization with a budget can operate at the frontier of technical excellence.