Software engineering with LLMs in 2025: reality check

Key Takeaways & Insights

Hype vs. Ground Reality: There’s a significant disconnect between the optimistic predictions of tech CEOs about AI’s role in software development and the day-to-day experiences of engineers. While executives tout AI tools as transformative, real-world adoption and efficacy are more nuanced.
Widespread, but Varied Adoption: AI coding tools are being rapidly adopted, especially within AI startups and big tech, but their effectiveness and integration vary widely across different organizations and use cases.
Step-Change in Productivity: Experienced engineers and industry veterans believe AI tools represent a fundamental shift in how software is built, comparable to the move from assembly to high-level languages. However, benefits are currently more evident at the individual developer level than at the organizational level.
Experimentation is Key: The most successful teams and engineers are those actively experimenting with AI tools, sharing tips, and iterating on what works.
Non-determinism as a New Challenge: Unlike past productivity leaps in programming, AI introduces non-determinism, requiring new approaches to verification and trust in code.
Long-Term Impact: The consensus among seasoned engineers is that AI will reshape software development, change cost dynamics (what’s easy or hard), and enable new kinds of ambitious projects.

Leverage MCP (Model Context Protocol): Adopt protocols like MCP to connect various tools, databases, and APIs, enabling conversational and agent-driven workflows.
Encourage Knowledge Sharing: Teams should create channels (e.g., Slack) to exchange AI tool tips, effective prompts, and use cases, accelerating collective learning.
Automate Repetitive Workflows: Identify and automate well-defined and repetitive tasks (ticketing, documentation, code reviews) using AI agents and integrations.
Iterative Experimentation: Regularly test and evaluate different AI tools for coding, code review, and documentation. Share findings and adapt workflows accordingly.
Focus on Use-case Fit: Apply AI coding tools to areas where they excel—such as generating first-pass solutions for well-defined tickets—while being cautious in novel, complex, or highly specialized domains.
Integrate AI in Existing Pipelines: For organizations with established APIs and modular services, integrating AI agents (via MCP or similar) can accelerate automation and productivity.

Adoption Rates: At AI-focused startups like Anthropic and Windinsurf, 90–95% of code is written with AI assistance; at Cursor, it's around 40–50%.
Big Tech Implementation: Google has integrated LLMs across many internal tools (autocomplete, code review, search), and is preparing for a 10x increase in code throughput. Amazon reports almost universal use of Amazon Q Developer Pro among internal devs, especially for AWS coding.
MCP Protocol: Open-sourced by Anthropic, adopted by OpenAI, Google, and Microsoft within months; thousands of servers now in use.
Case Study: Incident.io, a startup, found AI agents effective for well-defined tickets and shared practical prompting techniques internally.
Adoption Survey: A DX survey of 38,000 developers found about 50% use AI coding tools weekly; in leading companies, this rises to 60%.
Veteran Perspectives: Kent Beck (52 years in programming) finds AI tools have made programming more enjoyable and accessible, likening the change to the advent of microprocessors, the internet, and smartphones.

Overreliance on Hype: Don’t assume AI coding tools are universally transformative just because of executive claims; actual utility varies.
Blind Trust in AI Output: Reviewing and validating AI-generated code is still crucial, especially for novel or critical codebases, as hallucinations and errors persist.
Poor Fit for Novel Domains: AI tools often underperform in highly specialized or cutting-edge areas (e.g., novel biotech pipelines), where human expertise and context are irreplaceable.
Neglecting Organizational Fit: AI tools often work better for individuals or small teams than at the org level. Rolling them out organization-wide without a clear strategy can lead to disappointment.
Ignoring Feedback Loops: Not iterating or adapting based on AI tool failures or successes can stifle potential productivity gains.

Protocols & Tools: Explore MCP (Model Context Protocol), Amazon Q Developer Pro, Google Notebook LM, and other AI-integrated dev tools.
Further Reading: Articles by Armen Ronacher, Peter Steinberger, Simon Willison, and Bri Brigita for practical insights and experiences.
Surveys & Benchmarks: Consult DX’s developer surveys for current adoption benchmarks.
Experimentation: Start small—trial AI tools on well-defined tasks, measure impact, and scale successful practices.
Community Engagement: Participate in discussions, share experiences, and learn from ongoing experimentation in the developer community.
Keep Updated: Follow blogs and talks by leading engineers and AI tool creators for the latest developments and best practices.