Your team has spent a year with an AI assistant in the editor and has to decide whether to pay for coding agents, which open whole changes on their own. The natural expectation is another jump like the one in the first year. Two causal studies point to something more modest.
If your team already works with Copilot or something similar, one more agent will add little speed and will raise code complexity. In our experience, the jump comes with the system built around it: specifications and gates, automatic checks a change must pass to get in, that verify the agent's work.
What the two studies measure
Hao He, Shyam Agarwal, Bogdan Vasilescu and their co-authors compare, month by month, open-source GitHub projects that adopt an AI tool with similar projects that do not, to separate its effect from what would have happened anyway.
In the Cursor study (MSR 2026), speed rises sharply and briefly, while static analysis warnings (the problems a tool detects in code without running it) and complexity rise substantially and stay. The authors conclude that this increase explains much of the later slowdown.
The agents study, from January 2026, looks at agents that open pull requests (requests to merge a change into the shared code) on their own. The speed boost is large and comes early only when the agent is the project's first AI tool; with a prior AI-powered editor, the gain is minimal or short-lived. Quality gets worse in every case: about 18% more warnings and 39% more cognitive complexity, which measures how hard a fragment of code is to follow when you read it.
At onext we have seen it: most of our clients use Copilot and have a licence for some AI, and the change, in both speed and in the consistency and quality of what gets delivered, comes from building agents, skills and gates.
Limit Both studies look at open-source projects, and a company team has its own rules. The 18% and 39% figures come from the abstract of the agents study; the Cursor study gives no magnitudes. Neither measures the effect of specifications or gates.
What to do before paying for more agent
Measuring complexity is fine, but it is not the first step. First you have to decide what “done well” means for the team: a specification the agent has to follow and gates it cannot skip. Otherwise you are measuring the complexity of something nobody has defined. System first, counter second.
Manel Marcos, onext’s technical lead and co-founder, is clear about a complexity limit: “A complexity limit is only useful if it lives in the pipeline. If the agent can skip it with a comment or by switching off a rule, it is not a limit, it is a suggestion. What works is treating it like a test: a change that exceeds it does not get in, whether a person or an agent wrote it, and whoever wants to raise the threshold has to propose it and justify it like any other architecture decision.”
With that defined, measure complexity and static analysis warnings before switching the agent on, so you have something to compare with, and watch the trend every month alongside the delivery metrics you already track. If complexity keeps rising once speed has levelled off, the agent is charging in maintenance what it advanced in delivery.
If you want to see where your team stands before deciding, the self-assessment has 16 questions and is free (in Spanish).
Frequently asked questions
Does moving from an AI assistant to coding agents make a team faster?
According to a January 2026 study of open-source projects on GitHub, the speed boost is large and comes early only when the agent is the project's first AI tool. If the project already used an AI-powered editor, the gain is minimal or short-lived.
What happens to code quality when a team adopts agents?
In the same study, static analysis warnings rise by about 18% and cognitive complexity by about 39%, and the increase persists even when speed stops improving. The authors describe it as technical debt left by the agent.
What does the Cursor study say?
That adopting Cursor brings a large, statistically significant and transient increase in speed, and a substantial and persistent increase in static analysis warnings and code complexity. With a second analysis, the authors conclude that this increase explains much of the later slowdown. It was presented at MSR 2026.
What is cognitive complexity in code?
It is a metric created by SonarSource to measure how hard a method or function is to understand when you read it. Static analysis tools such as SonarQube calculate it, and in these studies it shows whether code is becoming harder to maintain.
What is a quality gate for a coding agent?
It is an automatic check that a change has to pass to get into the code, whether a person or an agent wrote it. One example is a complexity limit in the pipeline that rejects any change that exceeds it.
What can a team do before paying for coding agents?
Define what “done well” means with a specification the agent follows and gates it cannot skip, then measure the complexity and static analysis warnings of its code, and review the trend every month alongside delivery speed.
Sources
- Shyam Agarwal, Hao He and Bogdan Vasilescu, "AI IDEs or Autonomous Agents? Measuring the Impact of Coding Agents on Software Development", arXiv:2601.13597, v1 of 20 January 2026 and v2 of 27 January 2026.
- Hao He, Courtney Miller, Shyam Agarwal, Christian Kästner and Bogdan Vasilescu, "Speed at the Cost of Quality: How Cursor AI Increases Short-Term Velocity and Long-Term Complexity in Open-Source Projects", arXiv:2511.04427, v3 of 26 January 2026 (MSR 2026).
Bernat López is founder and CEO of onext, an AI boutique. He helps development and product teams work with AI with method —specification before building, a person who decides where there is risk and Spec-Driven Development— and applies to his own company what he proposes: onext runs on its own agentic system.
LinkedIn →