Short answer: Individual developers often feel faster with AI, but Google’s DORA 2024 research found that higher AI adoption was associated with small decreases in software delivery throughput and stability. The takeaway is not to avoid AI coding agents. It is to roll them out with the same delivery discipline—small batches, automated tests, review, and fast feedback—that DORA has long linked to high performance.
Most discussion of AI coding tools focuses on the individual: does the developer type less, ship faster, feel more productive? That is the wrong unit for an engineering leader. What matters for a team is the delivery system—how reliably valuable change reaches production—and the evidence there is more sobering.
The counterintuitive finding
Google’s DORA 2024 report (the Accelerate State of DevOps research) found genuine individual benefits from AI: higher self-reported productivity, more flow, and greater job satisfaction. But at the system level it estimated that a 25% increase in AI adoption was associated with roughly a 1.5% decrease in delivery throughput and about a 7.2% decrease in delivery stability.
In other words, the same adoption that makes individuals feel better can coincide with software that ships slightly slower and breaks slightly more often. Those are small percentages, but they point in the opposite direction from the marketing promise.
Why more code can mean less stable delivery
The mechanism is not mysterious. AI lets developers produce more change, faster—but throughput and stability depend on the whole pipeline, not on generation speed.
- Larger or more frequent changes strain review capacity and increase the chance a defect slips through.
- Review becomes the bottleneck. Generated code still needs a human to understand it; volume shifts load onto reviewers.
- Rework and churn rise when code is accepted quickly and corrected later.
- Verification is undercounted. METR’s 2025 randomized trial showed developers can believe they were sped up (by about 20%) while actually being slowed (by about 19%)—a perception gap that hides the true cost of validating output.
More generated code is not the same as more accepted, stable change.
Roll out agents as a delivery-system change
The productive response is to treat adopting AI coding agents as a change to the delivery system, governed by the practices DORA has repeatedly associated with high performance.
- Small batches. Keep agent-produced changes small and reviewable; resist large, unattributed diffs.
- Automated testing. Require tests to gate agent output, so speed does not erode correctness.
- Explicit review gates. A human approves high-impact changes; the agent does more of the work, not less of the oversight.
- Fast feedback and observability. Detect and revert regressions quickly; watch stability, not just volume.
- Progressive rollout. Expand agent autonomy as measured stability holds, not on a fixed schedule.
Metrics to watch
Pair the classic DORA four keys with AI-specific signals so a speed gain cannot hide a stability loss.
| Dimension | Metric | Why it matters with agents |
|---|---|---|
| Throughput | Deployment frequency, lead time | Detects whether faster generation actually reaches production |
| Stability | Change failure rate, time to restore | Catches the regression DORA associated with AI adoption |
| Review | Reviewer minutes per accepted change | Reveals load shifted from author to reviewer |
| Rework | Post-merge churn on agent changes | Flags “accept now, fix later” costs |
Where a managed platform fits
If agent work is scattered across individual laptops, none of these metrics are observable and none of these gates are enforceable. That is the argument for a managed layer.
MonkeyCode’s public materials describe exactly this shape: server-side environments, task and project history, and a review-oriented workflow rather than an in-editor autocomplete tool. For an engineering leader, the value is not that the agent is faster in isolation—it is that tasks, environments, and review evidence are recorded where throughput and stability can actually be governed. Compare that operating model against alternatives on the workflow comparison page and the self-hosting field guide.
Bottom line
AI coding agents can help individuals and still, at team scale, shave points off delivery throughput and stability if adopted without discipline. DORA’s data is a reminder that engineering performance is a property of the system, not of one developer’s typing speed. Roll agents out in small batches, behind tests and review gates, and measure the DORA keys as you go. The teams that treat AI as a delivery-system change—not a personal accelerator—are the ones positioned to keep both speed and stability.
Source boundary: Throughput and stability estimates are from Google’s DORA 2024 research and describe associations, not proven causation for every team. The perception-versus-reality figures are METR’s reported findings for early-2025 tools among experienced open-source developers, which METR marks as historical. All sources were checked July 20, 2026.