2025 Year End Reflections

2025 didn’t feel like progress — it felt like a phase change.

AI integration into daily work and life has deepened dramatically this past year. Across the industry, success stories are emerging from every domain. But beyond the headlines, what struck me most was how fundamentally it changed my own workflow.

Here’s what I learned.

Software Engineering

This was the year typing code became history. Code agents became the new norm for producing features and PRs. Throughout 2025, models and agents advanced significantly — from handling tiny snippets to completing small features without human intervention.

Challenges remain, especially for large-scale systems with strict quality requirements. But AI has become an essential tool for developers. At this point, I can’t imagine working without it.

At Work

For the past several years, despite growing experience, my output remained flat — roughly the same amount of code changes since 2018. This year broke through that plateau: 25% more results, with only two-thirds the time spent in development.

The secret? Multi-tasking. I now work on 3-4 tasks simultaneously, keeping either the GPU (LLM generation) or CPU (compilation and execution) constantly busy. The goal is simple: always have one running to maximize efficiency.

Wish list: trade more compute for better results. In enterprise settings, agent cost is rarely the constraint — human labor is orders of magnitude more expensive. Given that asymmetry, I want deep-research-like agents for every exploration during a task. Think of it like the reasoning effort dial in the GPT family: for critical code paths, crank it to extra high; for boilerplate or low-risk changes, medium is fine. Let me allocate compute where it matters most.

Side Projects

Outside of work, I made over 4,000 commits to personal GitHub repositories this year — yet many are small and incremental, driven by looping coding agents. This let me experiment with areas I’d never have time for otherwise: agentic orchestrators, MCP servers, VSCode Extension, CUDA, reinforcement learning, database internals, and performance optimization.

I maintain four subscriptions: Cursor (500 requests/month), Claude (switching between Max and Pro depending on workload), ChatGPT/Codex, and GitHub Copilot (1000 requests/month). I rotate between them to work around usage limits and explore new features after their major updates. Since most have weekly/monthly caps, leaving quota unused feels like waste — so I push to use everything available. These limits became a forcing function, driving me to write far more code than I otherwise would.

Wish list: harness fragmented time for meaningful contribution. Much of my day is spent between tasks — moments that feel wasted. I’d love to interact with coding agents through casual conversation during these gaps. Describe a feature idea, refine the approach through dialogue, and find a working prototype when I return to my desk. The pieces are almost there; the integration and user journey just need to catch up.

Challenges

Context remains the central challenge. The more precise context you provide, the better the results — but gathering and curating that context takes effort. Graphrag, Deepwiki etc. are helping to some extent, but still not enough.

Human attention is now the bottleneck. As AI handles more coding, the work shifts to steering and review. But effective review requires judgment built through practice — and junior engineers who leapfrog that process may struggle to catch AI’s mistakes. For teams, this is a real concern: how do you grow reviewers who never wrote the code themselves?

One of my side projects is a vscode extension that aims to tackle these 2 challenges - Markdown Review & Workspace Shortcuts . It helps with context by organizing files into shortcut groups (e.g. a feature in development). And it helps with review by enabling structured inline annotations that can generate AI prompts — a workflow that keeps humans in the loop while building judgment.

Looking forward, I hope to see AI/Agent become more autonomous and self-directed — where human attention is no longer the limiting factor.

Learning

Learning has become dramatically more efficient. Deep research tools help me digest papers and open-source codebases. Summarization handles YouTube videos and courses. I can absorb material much faster than before.

More importantly, coding agents make experimentation accessible. I can now build and test ideas in domains where I have minimal expertise — something that was simply out of reach before.

Life

AI integration into daily life has become surprisingly practical. I consult ChatGPT for health/finance questions and parenting advice, often with genuinely useful results. AI-powered tools help with product research, shopping decisions, and even sports improvement through video analysis.

These aren’t revolutionary applications — but they accelerate routine tasks and free up time for what matters.

Summary

If 2024 was about can AI do this?, 2025 was about why would I do this without AI?

The tools are no longer tools — they’re collaborators. The bottleneck is no longer capability — it’s attention. And the skill that matters most isn’t prompting — it’s agency and taste.

In 2026, I’m looking for the edges — the places where human and machine choose not to meet.