BREAKING
Vercel Resolves OpenTelemetry Blind Spot in AI SDK With New Inference Duration Histogram ● YBUSDT Approaches Potential Turning Point Amid Signs of Returning Buying Interest ● SEC to Resume Processing Registration Applications ● Rarible Unveils Rebuilt Cross-Chain Marketplace Featuring High-Speed Trading and New Rewards System ● Niko Bellic Appears in GTA 6: Voice Actor Michael Hollick Returns for Late-Game Cameo ● Ant International’s Alipay+ Connects Over 50 Global E-Wallets to Brazil’s Pix to Revolutionize Cross-Border Payments ● Federal Investigation Links Ransomware Negotiation Executives to ShinyHunters Hacking Group and FBI Data Breach ● White House National Cyber Director Sean Cairncross to Step Down Amid Transitioning Political Landscape and Unexplained Tech Outages ●
Blockchain Technology & Web3

The Great AI Coding Bottleneck: Why Agent-Written Pull Requests Wait Five Times Longer for Review

Photo: Documentation / News Illustration

As enterprise investment in generative artificial intelligence accelerates at an unprecedented pace, a paradoxical bottleneck is reshaping software engineering departments worldwide. Developers leveraging AI assistants can now generate and open pull requests in mere minutes, yet these automated contributions frequently stall in sprawling review queues, waiting for human engineers to step in. Industry data highlights a stark friction point in the modern software delivery lifecycle: agent-written pull requests wait an average of 17.6 hours for their first review, compared to just 3.4 hours for human-authored code. This widening chasm persists even as global enterprise spending on AI coding tools reached $4.0 billion, surging to seven times the volume recorded the previous year.

To examine how software development organizations are navigating this operational bottleneck, Behind the Startup sat down with Ori Keren, co-founder and CEO of LinearB. A veteran engineer who began writing BASIC code on a Sinclair Spectrum in 1987 before managing engineering groups at Interwise and CloudLock, Keren launched LinearB in 2019 alongside co-founder Dan Lines. The company originally set out to solve a persistent industry blindness—the reliance on gut feelings and status meetings to track software delivery—by connecting data from Git, Jira, and CI systems to give engineering leaders empirical insights. Today, as automated coding agents permeate the enterprise, that foundational mission has evolved into building a comprehensive context layer for both human developers and autonomous systems.

AI Made Writing Code Cheap. Getting It Reviewed Now Takes 5x Longer, Says LinearB’s CEO

The Shift from Code Generation to Delivery Bottlenecks

LinearB’s recent industry benchmarks, drawing from an analysis of 8.1 million pull requests across 4,800 teams, reveal two simultaneous realities shaping the current enterprise landscape. First, while generative AI has made writing code remarkably cheap, it has failed to make the overall delivery pipeline any faster. AI-assisted pull requests are roughly 2.6 times larger than unassisted ones, compounding the cognitive load required for code reviews. As autonomous agents generate higher volumes of larger, more complex contributions, the operational bottleneck

has shifted decisively downstream into the review, merge, and ownership phases.

AI Made Writing Code Cheap. Getting It Reviewed Now Takes 5x Longer, Says LinearB’s CEO

The second core challenge lies in the inherent limitations of autonomous coding agents. An agent typically approaches every assigned task in isolation, possessing a strong comprehension of local code syntax but lacking critical institutional and business context. An agent does not inherently know that a particular microservice experienced multiple production incidents during the previous month, that a specific file has undergone five complete rewrites within a single quarter, or which engineering team holds ultimate ownership of the service. Without access to this interconnected history—spanning from initial ticket creation down to final production release—agents struggle to produce code that aligns seamlessly with overarching architectural and operational standards.

Recognizing that traditional dashboards are fundamentally incapable of supplying this necessary operational context to autonomous agents, LinearB underwent a major structural evolution, rebuilding its core platform around a sophisticated knowledge graph of the entire software delivery lifecycle. As enterprise organizations scale their AI adoption, moving from isolated deployments of GitHub Copilot to multi-tool ecosystems incorporating platforms like Cursor and Claude Code, executive leadership frequently runs into a familiar wall: an inability to verify code quality, trace productivity gains, or accurately calculate total operational costs.

AI Made Writing Code Cheap. Getting It Reviewed Now Takes 5x Longer, Says LinearB’s CEO

Proving Return on Investment to the Boardroom

With boards of directors and chief financial officers increasingly demanding direct proof that multi-million-dollar AI investments are generating tangible returns, the dynamics of enterprise software procurement have shifted. While senior vice presidents of engineering and chief technology officers remain the primary executive champions, corporate finance departments are entering the evaluation room with growing frequency. Engineering leaders are under mounting pressure to present rigorous, quantitative metrics that can withstand the scrutiny of a CFO, moving far beyond the vanity metrics often promoted by tool vendors.

This financial scrutiny has fundamentally altered how technology contracts are won and renewed. Rather than relying solely on high-level cycle-time improvements, software platforms must now demonstrate immediate economic clarity during the evaluation phase. By connecting disparate delivery data, platforms can surface which internal teams are successfully deriving value from AI coding assistants, which teams are experiencing friction, and what the exact financial cost is for every delivered change.

AI Made Writing Code Cheap. Getting It Reviewed Now Takes 5x Longer, Says LinearB’s CEO

Token economics have introduced an additional layer of urgency to these financial conversations. According to LinearB’s benchmarks covering 253 engineering organizations, heavy AI users at the 90th percentile now spend $35.20 on model tokens per coding day, representing a nearly 75 percent increase over a three-month window. When executive leadership reviews these expanding operational line items, the demand for precise performance accounting intensifies. Organizations like Kraken have leveraged continuous integration context layers to diagnose review delays, identifying time zones and reviewer availability as primary constraints rather than code size, ultimately cutting their slowest review times by more than half within four months.

Community-Led Growth and the Evolution of Enterprise Sales

LinearB’s path to enterprise prominence highlights a hybrid go-to-market motion that pairs developer-focused community building with structured enterprise sales. Long before industry analysts established formal categories for developer productivity insights, the company cultivated an engaged technical audience through initiatives like the Dev Interrupted podcast and free developer tools. Originally conceived as a forum for engineering leaders to discuss the complexities of managing software teams, the podcast and associated community resources established deep industry trust.

AI Made Writing Code Cheap. Getting It Reviewed Now Takes 5x Longer, Says LinearB’s CEO

This organic familiarity often begins at the practitioner level, with individual staff engineers and engineering directors utilizing free tools or referencing open benchmarks for team planning. When these technical champions eventually advocate for organization-wide rollouts, executive leadership is already acquainted with the platform’s methodology. Enterprise sales teams then guide the prospective customer through rigorous security reviews, procurement procedures, and formal return-on-investment analyses. Consequently, the company’s largest enterprise deals frequently trace back to long-standing community relationships established well before the formal sales cycle commenced.

Rethinking Metrics for the Age of Autonomous Agents

As software engineering organizations look toward the coming years, industry leaders emphasize the urgent need to abandon outdated productivity metrics. Traditional indicators such as raw commit counts, pull request throughput, and AI adoption percentages—such as reporting that ninety percent of a workforce utilizes an assistant—fail to provide meaningful insight into business value or code quality. Instead, enterprise leadership is encouraged to focus on the cost per effective delivered change. This metric divides total personnel and token expenditures by the number of changes that successfully merged, shipped to production, and remained stable without requiring immediate rework.

AI Made Writing Code Cheap. Getting It Reviewed Now Takes 5x Longer, Says LinearB’s CEO

By tracking effective changes alongside model-specific token costs, engineering organizations can identify significant operational variances across teams and identify optimal development workflows. As autonomous agents take on an increasingly prominent role in writing code, the software delivery lifecycle will continue to demand sophisticated context layers, governance frameworks, and spend optimization tools. For engineering executives preparing for upcoming budget and board reviews, shifting accounting practices toward the true cost of effective delivery remains a critical step in navigating the transition toward agent-driven software development.

What is your reaction to this story?

Raul Delapena Setiawan

Writer and contributor at Bitcoin 98.

Comments (0)