The engineering landscape is undergoing a seismic shift. As AI-powered development tools proliferate and reshape how teams build software, engineering leaders face a paradoxical reality: productivity may decline before it improves. This counterintuitive prediction emerged during a recent conversation with Ori Keren, CEO of LinearB, and Dharmesh Thakker, General Partner at Battery Ventures, who gathered with engineering leaders in San Francisco to explore what the next 12 to 24 months hold for developer productivity.
The discussion revealed a nuanced picture of transformation, one where experimentation precedes optimization, where senior developers become more valuable rather than less, and where the definition of productivity itself is evolving from lines of code to business impact.
AI will slow developer productivity before it accelerates it
The most provocative insight of the evening came from Keren's opening prediction. He noted that despite the many levers available to increase productivity, teams are still in a norming and storming phase. Because of this, Keren expects developer productivity to actually decline in 2025 as organizations work through a challenging period of experimentation, before peaking in 2026 and 2027.
This prediction challenges the prevailing narrative that AI tools will immediately unlock massive productivity gains. Instead, 2025 will be a year of process adjustment, a necessary investment period before organizations can realize sustainable improvements. The reasoning is sound. Teams are still learning how to integrate AI into their workflows, adjusting processes, and discovering which use cases deliver genuine value versus which create new bottlenecks. Every technological shift follows this pattern of initial disruption followed by normalization and eventual optimization.
Thakker reframed the productivity conversation entirely, moving beyond throughput metrics. He argued that the goal is not measuring lines of code, but measuring how teams impact the business by combining application logic with foundational models to build killer applications at record speed.
This shift from measuring output to measuring outcomes represents a fundamental evolution in how engineering leaders should think about productivity. AI-enabled development increases code volume, but volume alone is meaningless if it doesn't translate to customer value. The real question becomes whether teams are shipping features that drive business impact faster.
The conversation also highlighted how AI is exposing downstream constraints. When developers can generate code faster, bottlenecks in code review, CI/CD pipelines, and rework become more visible and more painful. Organizations that fail to address these systemic issues will find their productivity gains limited by the slowest part of their delivery pipeline.
Audience participation reinforced that the clearest short-term wins are appearing in repetitive, operational work such as dependency maintenance, boilerplate generation, and internal knowledge triage. These low-risk, high-volume tasks are ideal candidates for automation, freeing developers to focus on higher-order problems that require human judgment and creativity.
Developer experience will separate teams that ship in minutes from teams stuck for weeks
Developer experience emerged as the critical enabler of productivity gains, with both panelists emphasizing that organizations must systematically remove workflow friction to capitalize on AI capabilities.
Keren explained that the most important step is establishing a framework to fix problems. Whether an organization collects metrics quantitatively, qualitatively, or both, simply gathering data is just the beginning. The emphasis on frameworks over metrics alone is crucial. Too many organizations collect data about developer experience, including build times, flaky tests, and review queue length, but fail to act on those insights. Measurement without action creates frustration rather than improvement.
Keren painted a stark picture of the divergence ahead. He believes that leaders with an awareness of developer experience who actively want to improve it will soon take their teams from an hour of commit-to-production time down to just minutes because they will have AI agents solving pipeline issues for them. Conversely, those who fall behind will remain stuck with cycle times measured in weeks.
This prediction suggests a widening gap between organizations that invest strategically in DevEx and those that do not. The competitive advantage will not come from adopting AI tools in isolation, but from building integrated systems where AI agents can identify and resolve pipeline issues, automate routine maintenance, and keep developers in a flow state.
Thakker offered a simple but powerful framework for prioritizing DevEx investments: take the highest-impact tasks and help developers do them more effectively, then take the lower-impact things that simply have to get done and automate them.
This two-pronged approach provides a clear lens for evaluating DevEx initiatives. The goal is not to eliminate all friction, but to ensure friction does not slow down the work that matters most. Future DevEx improvements were described as more unified workflows connecting coding, testing, deployment, and feedback loops more tightly. Documentation and knowledge sharing are expected to shift toward AI-generated and AI-maintained systems, reducing dependence on static internal wikis and making engineering knowledge easier to retrieve and apply when needed.
The investment case for DevEx was framed in ROI terms. Organizations that systematically remove workflow friction should see dramatic reductions in commit-to-production cycle time, creating a compounding advantage over teams that fail to modernize their development infrastructure.
Structured AI rollouts lift engineering velocity
The discussion of AI-assisted workflows revealed broad experimentation across autocomplete, code generation, and AI-assisted reviews. When asked how many in the audience were actively adopting or experimenting with at least one of these technologies, nearly every hand went up, a clear signal that these tools are rapidly becoming table stakes.
One particularly insightful challenge came from the audience regarding the idea that chat should remain the default interface for AI assistance. Natural language code interfaces are more valuable when embedded directly into editing, review, and knowledge workflows rather than isolated in standalone chat windows. This observation points toward a future where AI assistance is contextual and workflow-aware rather than conversational.
Thakker emphasized the economic logic driving adoption, noting that all the ancillary work developers never got a computer science degree to do can now be automated, which simply makes economic sense. Documentation, unit tests, syntax fixes, and dependency updates consume significant developer time but offer limited intellectual satisfaction. Automating them not only improves productivity but also enhances developer experience by freeing teams to focus on more fulfilling work.
Trust in AI output was described as conditional rather than absolute. Thakker observed that automating low-end pull requests is becoming standard practice, indicating that trust is not necessarily the issue, provided organizations remember to verify more complex changes.
This risk-based approach to automation makes sense. Low-risk pull requests, boilerplate work, and dependency updates are easier to automate with confidence, while complex or high-impact changes still require human verification. The key is having systems that can distinguish between these categories and route them appropriately.
Structured rollout practices were highlighted as critical for realizing measurable gains. Recent research from Google found that structured rollouts of AI tools yielded a 21 percent velocity improvement for enterprise-grade tasks, with senior developers seeing even larger gains. This suggests that organizations should not simply give developers access to AI tools and hope for the best. They need frameworks, training, and clear expectations to maximize adoption and impact.
Agentic AI will multiply senior developers once governance is in place
The conversation took a more speculative turn when discussing agentic AI, autonomous systems that can pick up queued tasks, open pull requests, fix bugs, and respond to SRE or security issues without human initiation.
Keren drew a clear distinction between assisted and agentic workflows, suggesting that agentic use cases are essentially a year behind current capabilities. While early experiments are visible, he noted that organizations do not yet have the confidence to deploy an agent that can independently take a task from a JIRA queue and issue a pull request.
This adoption lag exists despite the technical feasibility of many agentic use cases. The constraint is not capability, it is trust, governance, and operational readiness. Enterprise organizations need to solve thorny questions around IP ownership, model training boundaries, legal exposure, data handling, and compliance requirements like SOC 2 before they can confidently deploy autonomous agents in production environments.
Keren predicted the emergence of new organizational roles to support this shift, forecasting that companies will soon hire Heads of AI Development specifically to define and implement policies. This role would sit at the intersection of engineering leadership and governance, establishing guardrails for where autonomous behavior is permitted and how policies are enforced across engineering systems. It represents a recognition that agentic AI is both a technical challenge and an organizational design challenge.
The discussion also touched on the rise of citizen developers and low-code development as a likely outcome of better AI tooling. Product managers and other non-technical team members are increasingly able to build prototypes and internal automations without writing code. While this democratization has benefits, it also introduces risks around shadow IT, inconsistent practices, and technical debt.
One audience member shared their experience managing this tension, describing how they encourage senior developers to think more like product managers by expressing problems in terms of user value and business outcomes rather than implementation details. This shift in mindset helps developers leverage AI tools more effectively while also preparing them for a future where natural language interfaces may replace traditional coding for certain tasks.
The panel acknowledged that agentic AI will fundamentally change what it means to be a senior developer. In the future, a strong senior developer might function as a five-person team, activating agents to handle queued tasks, perform AI-based reviews, fix SRE problems, and address security issues. This massive multiplier effect would make senior developers dramatically more valuable, but also require new skills around orchestrating and supervising autonomous systems.
Engineering leaders who modernize now will pull far ahead
The conversation revealed several critical insights for engineering leaders navigating this transition. Productivity will dip before it improves, and organizations should set realistic expectations for 2025, viewing it as an investment year focused on experimentation and process refinement rather than immediate productivity gains.
Engineering leaders must redefine productivity around business impact. Move beyond measuring lines of code or pull requests merged. Focus on customer value delivered, architectural quality improved, and time-to-feedback reduced. Investing strategically in developer experience is essential. Organizations that systematically remove workflow friction will create compounding advantages. Start with measurement, but build frameworks and infrastructure to address the actual problems you identify.
Senior developers will become more valuable, not less. Experienced engineers are better positioned to validate AI output, make higher-order design decisions, and connect technical changes to product outcomes. Leaders should invest in helping them adopt AI tools through structured rollouts and clear expectations.
Organizations must prepare for agentic AI, but should not rush it. While the technology is advancing rapidly, enterprise adoption will lag due to governance, compliance, and trust concerns. Start defining policies now for where autonomous behavior will be permitted. Embrace the rise of citizen developers, but manage the risks. Low-code and no-code tools will empower more people to build software, but without proper governance, this can create shadow IT and technical debt.
The future of developer productivity is not about replacing developers with AI. It is about fundamentally reshaping how teams work, what they optimize for, and where they focus their energy. Organizations that navigate this transition thoughtfully, with attention to both technology and culture, will build sustainable competitive advantages that compound over time.
To hear more of Ori Keren and Dharmesh Thakker's predictions for the future of developer productivity, listen to the full episode on the Dev Interrupted podcast.
