Probabilistic engineering and the 24-7 employee

Software is quietly becoming a probabilistic system. The deterministic contract—write, test, ship, know it works—is breaking. Inside top AI-native companies, codebases are becoming things you believe work, "with a probability you can no longer precisely state."

Davis describes his side project Compound Loop, orchestrating multiple frontier models to autonomously write, review, and merge code. He sets it running before bed and wakes to triage PRs that didn't exist the night before.

The 24-7 Employee

"For the first time in the history of knowledge work, the person who went home did not take the only copy of their brain with them."

The 24-7 employee isn't a person working around the clock—"it is a person whose agents work with enormous parallelization."

Role Splitting

Two opposing forces:

Upward shift: Top engineers become more effective PMs; top PMs become system architects; top architects think about distribution and market shape. "For this group… the work is more leveraged than it has ever been."

Downward fragmentation: Many engineers become "spec writers, reviewers, and agent babysitters"—translating intent into machine-readable prompts and grading outputs. "The 2026 equivalent of data entry, dressed up in new terminology."

The pay gap between the top tercile running fleets and the middle tier managing the exhaust will exceed the old engineer-vs-sales-rep gap.

Defensible moats: kernel performance, compiler design, hardware abstraction—areas still requiring determinism.

Jevons Paradox Applied to Code

Jevons (1865): more efficient steam engines led to more coal consumption. As code generation approaches zero marginal cost, teams write vastly more. "The companies that believe the scaling laws are unbounded are building accordingly, and they will be the power-law-distributed winners."

When supply explodes, "selection becomes everything." Direction, selection, and coherence become the highest-leverage work. "Production is not where the work gets hard anymore."

Probabilistic Engineering

Core asymmetry: "generation has become cheap, but validation has not." A 500-line PR in under a minute, but catching a subtle bug takes a senior engineer an hour. Review scales worse than linearly because context grows as more of the codebase is agent-written.

Failure mode: not dramatic collapse but "slow, silent degradation"—generation rises, review quality falls, unnoticed defects accumulate.

Industry Tiering

1. Deterministic tier – avionics, medical devices, financial trading, nuclear. "The cost of a silent correctness failure in a fly-by-wire system is not a customer complaint—it is lives."
2. Probabilistic tier – consumer software, SaaS, internal tools, content, marketing. "A probabilistic team willing to ship, measure, and correct can out-learn a deterministic competitor by an order of magnitude per quarter."
3. Convergence zone – insurance, healthcare, enterprise. Probabilistic methods creep upward "ten percent at a time."

Winners: "the teams that know which tier they are in, resist the temptation to pretend they are in a different one."

The Agentic Fleet

Davis rejects "factory shift" for "the agentic fleet"—but most operators run "a swarm of brittle contractors" rather than a well-drilled Navy. A fleet has composition, coordination, command structure, and watch shifts.

New workday: morning triage and merge, midday high-leverage work, afternoon review and redirection, evening handoff with overnight specs. "You wake up ahead of where you ended, provided the review discipline holds."

Build for the Model You Don't Have Yet

"the model we are using today is the dumbest model we will ever use." Frontier capability will "meaningfully exceed today's in the next six to twelve months." Organizations aren't building muscle for today's model but for 2027/2028. Irrelevancy "does not announce itself—it arrives as a gradual inability to keep up."

The Muscle We Will Lose

Referencing his prior essay ("What We Owe the Minds We Create," Part V): "if you never build, you lose the ability to evaluate what is being built."

Junior engineers leaning on AI since week one ship fast but "when it fails in a way the model did not anticipate, they often cannot find the bug." Key skills at risk: taste (not learned by clicking approve), judgment (not developed by accepting a machine's answer in five seconds), craft (not acquired by reviewing agents' work).

The apprenticeship model breaks when juniors ship through agents and seniors review agent output. "The current generation of senior engineers are the last cohort fully trained in this old methodology."

Prescription: "do it without the fleet… deliberately and regularly, the hard way, on something that matters. Keep the muscle."

The Uneasy Part

Davis does not resolve to optimism. Plausible cases include exhausted employees, fragmented unrewarded roles, a generation lacking craft, teams confusing output volume with quality, widening organizational gaps.

Closing: "Let us build organizations for the model we do not have yet… Let us dispatch the night fleet and sleep well knowing the work is underway—and stay awake to the possibility that some of what comes back is wrong in ways we are no longer trained to see."

References: Joint research with Proximal and Modular on testing frontier agentic systems (modular.com/blog/how-frontier-coding-agents-built-a-video-diffusion-pipeline-on-max).