For millions of web developers worldwide, Next.js stands as a foundational framework for building modern, high-performance web applications. However, the immense popularity of the ecosystem comes with a significant operational burden: managing the influx of bug reports, feature requests, and technical queries submitted to its public GitHub repository. With an average of 36 new reports arriving each week—a volume amplified significantly by the proliferation of AI-powered coding assistants that streamline the bug-filing process—the Next.js maintainer team faced an escalating triage crisis.
By January 2025, the open issue backlog reached an unsustainable peak of 3,109 reports. The sheer volume threatened to obscure critical current regressions beneath layers of fixed bugs, duplicate submissions, and tickets concerning deprecated or unsupported software versions. Traditional attempts to manage the overflow through simple time-based automation proved inadequate. Recognizing that manual review was no longer scalable, engineering teams at Vercel deployed a sophisticated autonomous agent framework. Within weeks, this artificial intelligence intervention successfully slashed the backlog below 1,000 issues while maintaining exceptional accuracy, offering a potential blueprint for open-source maintenance in the generative AI era.
The Genesis of the Triage Crisis and the Limits of Time-Based Automation
The modern software development lifecycle has been radically accelerated by AI tooling. While coding agents have made it remarkably easy for developers to generate detailed bug reports and comprehensive reproduction steps, this capability introduces downstream friction for repository maintainers. Every reported anomaly demands rigorous verification: maintainers must reproduce the reported bug, test it across multiple historical and current versions of Next.js, and comb through years of commit histories and pull requests to determine the appropriate resolution.
Prior to deploying AI-driven investigation tools, the Next.js team attempted to stem the tide using conventional repository management heuristics. In January 2025, as the backlog reached its historical zenith, maintainers implemented an automated workflow designed to flag issues left inactive for two years—a threshold later shortened to 18 months—and subsequently close them following a brief grace period.
While this crude mechanism helped reduce the active queue to 2,244 reports by August 10, 2026, inactivity proved to be a deeply flawed proxy for relevance. An aging issue ticket might represent an already resolved problem, a duplicate submission, expected framework behavior, or a legacy bug tied to obsolete dependencies. Conversely, it could also highlight a genuine, persistent regression that had simply evaded human attention. Because a static timestamp cannot distinguish between these nuanced scenarios, time-based automation inevitably resulted in the erroneous closure of legitimate bug reports. To resolve the backlog responsibly, Vercel required a system capable of conducting deep contextual research.
Deploying Eve and the Closability Agent
To move beyond blunt time-based heuristics, engineers built a specialized research agent named closability, powered by Vercel’s open-source agent framework known as eve. Designed specifically for complex software maintenance tasks, eve provides persistent execution sessions, scheduled automation routines, structured data outputs, and the ability to safely execute code within isolated environments via Vercel Sandbox infrastructure.
Operating within a secure, isolated sandbox pre-configured with the Next.js repository, Node.js, Playwright, and Chromium, the closability agent systematically investigates individual issues. The agent replicates the environment, runs test suites, checks code changes, and gathers verifiable evidence regarding the validity of a bug report.
Despite its advanced capabilities, the agent operates under strict operational boundaries. To safeguard the integrity of the repository, closability is strictly read-only outside of its designated sandbox environment. It lacks the administrative permissions required to publicly comment on issues, close tickets, push code commits, or trigger deployments. Furthermore, to mitigate security vulnerabilities such as prompt injection attacks—where malicious actors embed hidden instructions within issue descriptions or user comments—the agent is explicitly programmed to disregard extraneous directives found within untrusted issue text.
Executing the Backlog Reduction Campaign
The large-scale triage campaign commenced in earnest when the closability agent was unleashed against the complete historical backlog utilizing GPT-5.6 Luna with maximum reasoning configurations. Given that an exhaustive, high-fidelity investigation of a single issue required an average of 30 minutes of computational analysis, Vercel engineers gradually scaled concurrency until approximately 200 eve sessions operated simultaneously.
The agent compiled its findings into a structured Close Queue, which human maintainers subsequently audited. Rather than blind execution, engineers reviewed the evidentiary foundation compiled by the agent before finalizing any administrative actions. By September 4, 2026, just three weeks after the intensive campaign began, the repository had closed 1,462 issues—including adjustments made outside the automated review—successfully driving the total open backlog below the 1,000-mark threshold for the first time in years, even as 218 new reports arrived concurrently.
A granular breakdown of the closed issues illustrates the diverse nature of the historical backlog:
- Already fixed: 543 issues (37%)
- Duplicate submissions: 278 issues (19%)
- Expected behavior: 237 issues (16%)
- No longer reproducible: 89 issues (6%)
- Unsupported or obsolete versions: 66 issues (5%)
- Other categories: 249 issues (17%)
- Total closed: 1,462 issues (100%)
Ensuring Quality Control and Community Recourse
A critical concern in automating repository triage is the potential alienation of the developer community through false-positive closures. To maintain transparency and accountability, Vercel implemented preventative safety measures prior to launching the review. The team deployed a specialized GitHub Action designed to provide community members with a clear recourse window if an issue was closed erroneously.
Whenever a maintainer or agent closes an eligible issue, the automated bot posts a public comment notifying the author. The system grants a strict 14-day grace period during which community members can request a reopening by replying with a specific command and a justification. If the original author or commenter responds within the window, the issue is reopened automatically.
This feedback mechanism served as a vital empirical test of the AI’s decision-making accuracy. As of September 4, 2026, out of 1,462 closed issues, only 3 had been successfully reopened by the community, translating to a remarkable 99.8% retention rate for the closed queue. Industry analysts note that such high precision underscores the effectiveness of pairing autonomous AI research agents with human-in-the-loop oversight.
The Broader Maintainer Agent Ecosystem and Future Outlook
The closability agent represents only a single component of a broader suite of independent eve agents collectively designated as the "Maintainer Agent." Within this unified architecture, specialized companion agents handle distinct phases of the software lifecycle, including reproducing complex bugs, verifying reported anomalies against canary releases, performing automated bisect operations to isolate introducing commits, generating end-to-end test suites, and preparing initial patch fixes.
These agents are integrated into a centralized Next.js operational dashboard that visualizes backlog trends, real-time research outputs, and triage queues. When multi-stage investigations are required, the dashboard orchestrates the handoff between agents, records execution telemetry, and dispatches high-priority diagnostic findings and daily summaries directly to internal team communication channels via Slack.
To prevent the backlog from regressing to its historical highs, Vercel has instituted a continuous, automated maintenance schedule. Every Monday, the closability agent automatically investigates up to 100 open issues that have experienced no activity for at least 30 days, prioritizing tickets that have never undergone automated review. Furthermore, for the most unambiguous categories of closure, maintainers have authorized the agent to autonomously resolve up to 25 clear-cut cases per week without mandatory human sign-off, though all code modifications continue to require rigorous human peer review before integration into the framework codebase.
Implications for Open-Source Sustainability
The successful reduction of the Next.js backlog highlights a transformative shift in how massive open-source projects can scale their operational capacity. As global software development accelerates and the volume of community contributions surges, human maintainers face cognitive and temporal bottlenecks that threaten project health.
By strategically deploying constrained, sandbox-isolated autonomous agents to handle the labor-intensive mechanics of research and verification, maintainers can reclaim thousands of hours of engineering time. Vercel’s initiative demonstrates that artificial intelligence, when deployed with robust guardrails, conservative confidence scoring, and transparent community feedback loops, can effectively manage administrative bloat without compromising the collaborative spirit and responsiveness essential to open-source software ecosystems.
