The Great Productivity Paradox: More Code, Same Velocity

The Great Productivity Paradox: More Code, Same Velocity

The data from 2025-2026 reveals a fascinating paradox. Individual developers are writing dramatically more code, but project delivery hasn't accelerated proportionally.

Faros AI's analysis of over 10,000 developers across 1,255 teams found that high AI adoption teams completed 21% more tasks and merged 98% more pull requests. But here's the catch: PR review time increased by 91% [3]. The pressure simply shifted upstream to verification and decision-making.

Boris Cherny exemplifies the individual productivity explosion. He's submitting 200 AI-written pull requests per month without even using an IDE, orchestrating development entirely through AI agents [2]. Malte Ubl built major open-source projects in record time using similar approaches. These aren't isolated cases—they're previews of a new development paradigm.

Yet Agoda's experience shows why individual gains don't automatically translate to organizational velocity. Despite AI dramatically increasing individual output, their overall project delivery saw only modest improvements [3]. The constraint moved from "how fast can we write code?" to "how well can we specify what we want and verify what we get?"

From Coding Bottlenecks to Context Gaps

The real constraint in 2026 isn't code generation—it's context transfer. Greg Foster, CTO at Graphite, puts it bluntly: "Context is the real constraint over security or quality" [6].

This manifests in several ways. Tacit knowledge—the unwritten rules, micro-decisions, and institutional memory that guides development—remains stubbornly difficult to feed into AI systems. Developers spend increasing time extracting context from AI-generated code, often leading to technical debt when shortcuts are taken [6].

A 450-ticket experiment using Codex and Claude revealed the judgment gap clearly. The agents excelled at execution but consistently failed at decision-making. They amplified unglamorous issues like test cleanup and unnecessary complexity because they lacked the judgment to prioritize or set boundaries [4]. As the researchers noted: "Agents execute extremely well. They don't choose which decisions to make."

The solution emerging from successful teams is what Agoda calls the "grey box" approach: provide precise specifications upfront, then rigorously verify results. It's specification-driven development on steroids, where the quality of your specs directly determines your velocity.

The Rise of Developer-Orchestrators

The developer role is fundamentally refactoring. GitHub's analysis suggests developers are becoming "orchestrators of AI-driven ecosystems" rather than hands-on coders [5]. This isn't just a buzzword—it's a practical shift with real implications.

Successful AI-assisted developers now spend their time on:

• Defining intent and system boundaries
• Crafting effective prompts and iteration cycles
• Reviewing and validating AI output
• Making architectural decisions
• Explaining existing codebases to AI systems

Andrej Karpathy captured this transition: "The profession is being dramatically refactored... developers are becoming 10X more powerful" [2]. But that power comes with new skill requirements.

The overlap with product management is real and intentional. As Dario Amodei predicts AI will write 90%+ of code, developers who thrive will be those who can bridge technical execution with product judgment [2]. The value increasingly lies in tech lead capabilities: understanding user needs, making trade-offs, and maintaining system coherence.

Nordic Lessons: Iteration Over Perfection

From our Nordic perspective, the most successful AI-assisted development approaches embrace iteration over perfection—a principle deeply embedded in Nordic design and engineering culture.

The teams seeing genuine velocity gains follow predictable patterns:

• They start with smaller, more aligned teams where context transfer is easier
• They prioritize rapid iteration cycles over comprehensive upfront planning
• They treat AI output as a first draft, not a final product
• They invest heavily in verification infrastructure

GitHub's Copilot Spaces represents this philosophy in practice—providing AI with bounded context rather than trying to feed it everything [5]. The constraint becomes a feature, forcing clearer thinking about what the AI actually needs to know.

The "swarm direction" approach is particularly Nordic in character. Rather than trying to perfectly specify everything upfront (a very American engineering impulse), successful teams provide clear boundaries and let AI agents explore within those constraints. It's controlled emergence rather than rigid planning.

Practical Frameworks for the Judgment Economy

For builders navigating this transition, several frameworks are proving essential:

Spec-First Development: Write specifications as if you're briefing a very capable but context-free contractor. Agoda's shift to treating developers as "Solution Architects" with specs as primary deliverables isn't just organizational restructuring—it's adapting to AI's strengths and limitations [3].

Agent Orchestration: Master the art of chaining AI agents for complex workflows while maintaining human oversight at decision points. The key insight from the 450-ticket experiment is that full automation is fragile, but interactive approaches with clear human boundaries are powerful [4].

Verification Infrastructure: Build robust testing and review processes that can handle the 98% increase in code output. The teams succeeding aren't just generating more code—they're processing it more effectively.

Context Architecture: Develop systematic approaches to feeding AI the right context. This means better documentation, clearer architectural decisions, and explicit knowledge management practices.

The Bigger Shift: When AI Builds the Software

The post-code era isn't about AI replacing developers—it's about fundamentally changing what software development means. When code generation becomes commoditized, judgment becomes the scarce resource.

This creates new categories of competitive advantage. Companies that can specify what they want clearly and verify what they get accurately will move faster than those with better individual programmers. The constraint shifts from technical skill to product clarity and organizational alignment.

The Nordic countries are well-positioned for this transition. The cultural emphasis on clear communication, iterative design, and pragmatic problem-solving aligns naturally with AI-assisted development patterns. The challenge will be scaling these approaches beyond small, aligned teams.

For Up North AI and similar organizations, this represents a massive opportunity. As code becomes free, the premium goes to those who can orchestrate AI effectively, make sound technical judgments, and maintain system coherence at scale. The judgment economy rewards sharp thinking over fast typing.

The transformation is already underway. The question isn't whether AI will change software development—it's whether you'll adapt your processes, skills, and organizations to thrive in a world where code is abundant but good judgment remains scarce.

Sources

1. https://metr.org/blog/2026-02-24-uplift-update
2. https://newsletter.pragmaticengineer.com/p/when-ai-writes-almost-all-code-what
3. https://www.infoq.com/news/2026/03/agoda-ai-code-bottleneck/
4. https://medium.com/@peterohsw/coding-agents-cant-automate-away-human-judgment-f4ebe1baa35e
5. https://github.blog/ai-and-ml/the-developer-role-is-evolving-heres-how-to-stay-ahead
6. https://thenewstack.io/context-is-ai-codings-real-bottleneck-in-2026
7. https://hai.stanford.edu/ai-index/2025-ai-index-report
8. https://arxiv.org/abs/2603.27438