The standard model of AI coding tools is a rising line. Models improve. Context windows grow. Agents learn your codebase. The ceiling — the limit of what you can safely delegate — goes up.
Nine independent datasets measured what happens to the systems that delegation depends on: the specifications, the review processes, the judgment. All three degrade over time. The ceiling doesn't rise. It sinks.
The Spec
The delegation ceiling is set by what you can specify, not by what the model can generate. That distinction matters because generation capability is what improves. Specification capability is what degrades.
Chakrabarti tracked 247,694 instruction lifetimes across 1,867 repositories. CLAUDE.md files — the primary agent specification mechanism — grew 226% over their lifetimes, adding 4.9 net instructions per commit. Older instructions became exponentially stickier: deletion hazard dropped 3.2% per commit of age. Adding a line is cheap. Removing one requires reasoning through every downstream behavior it might affect.
Denisov-Blanch et al. (ASE '26) measured the response across 441 repositories: 73.8% of AI configuration artifacts were committed once and never modified. Set-and-forget. The teams that maintain specifications face unbounded growth. The teams that don't pay a quality tax — cognitive complexity nearly doubles (+53% without maintained config versus +27% with it).
Specification debt is the mirror of technical debt — except it accumulates in instructions that nobody tests.
The Gate
Zhong et al. analyzed 1.02 million reviewed pull requests across 207 projects spanning three eras: human-centric, LLM-assisted, and agentic. Agent-involved reviews were faster. They were not better. Agent-authored PRs passed through with less discussion, less scrutiny, faster merges.
The 3,100 Opinions study built a 26-construct causal model from 38,709 documents and named the structural role: code review is "the control point through which a coding agent's effect on software is decided." The AI doesn't determine whether delegation helps or hurts. The team does, through review.
The control point weakens under exactly the throughput pressure that delegation creates. Orrù et al. tracked 182 repositories longitudinally post-merge. Agentic code showed comparable overall maintenance but significantly higher corrective maintenance and more security issues. The sharpest number: each 10 percentage-point increase in no-review rate produced approximately 6% more maintenance burden.
Review is the gate. Delegation widens it. The cost shows up downstream, in corrective maintenance that the generating team doesn't pay.
The Judgment
Data & Society's "Oversight Fallacy" names the third mechanism: "When routine tasks are delegated away, users may lose access to the repetition through which judgment develops." Selective attention depends on judgment. Judgment depends on repeated practice. Delegation removes the practice.
Park et al. observed 19 experienced developers managing AI coding agents. They concentrate effort in planning — and delegate supervisory work to other agents. Supervision itself gets delegated. The loop is recursive.
MAST (NeurIPS 2025, validated 2026) analyzed 1,642 execution traces across seven multi-agent frameworks: 79% of failures are specification problems. The failure is in what humans specify, not in what models compute.
The Direction
| Stage | Mechanism | Evidence | Direction |
|---|---|---|---|
| Pre-generation | Specification debt | 226% growth, 73.8% never updated | ↓ |
| At the gate | Review degradation | 1.02M PRs: faster, not better. 10pp → 6% burden | ↓ |
| Post-delegation | Judgment erosion | 79% spec failures. Supervision delegated recursively | ↓ |
Three mechanisms, all independent, all self-reinforcing. Specifications degrade because maintenance costs exceed accumulation costs and no one budgets for the difference. Review degrades because the throughput that justifies delegation pressures the gate that makes delegation safe. Judgment erodes because delegation removes the repetition that builds the capacity to oversee delegation.
None of these are addressed by improving the model. A better model generates more code, which increases throughput pressure on review, which widens the gate, which defers more costs downstream. Model improvement makes the first two mechanisms worse. The third is orthogonal — it operates on the human, not the tool.
The ceiling is not set by what the model can do. It is set by what the human system can specify, review, and maintain judgment about. That is what sinks.