A Year Into Agentic Coding, a Developer Says AI Now Writes Most of His Code but Still Falls Short as a Writer or Teacher

How much of a developer's code does AI write now?

A year after he last tallied how he uses AI, a research engineer at Salesforce AI Research says his habits have changed more than he expected. The biggest shift is in his job. He now spends most of his day directing coding agents, and he says machines write more than 90 percent of his code.

He previously worked as an engineer on Microsoft's Windows AI team. At Salesforce, he spends more time supporting scientists and building rapid prototypes than managing a large team of engineers. He estimates he spends about $3,000 a month on tokens, the units of text that AI models process and that vendors charge for.

Coding has changed the most

Last year, he wrote about how GitHub Copilot's autocomplete transformed his coding. That feature still works well, he says, but it has become obsolete for him. These days he typically runs five or six agents, including Claude Code, OpenCode, Codex and Pi. He hands each one a feature, a bug or a chore, and each agent works in its own copy of the codebase, writes code, runs tests and even opens pull requests.

The speed is the main draw. He says agents can prototype a new feature in about 20 minutes, and he has used them to test audio code on macOS and iOS in minutes. He has also had agents diagnose and fix urgent bugs from a short description while he worked on other things.

Not everything goes smoothly. He says agents left to run on their own produce code that is poorly functioning, over-architected and hard to read. He recently replaced about 4,000 lines of code produced by an agent with a few hundred lines he wrote by hand, and he says the result was more flexible, easier to test and more general. Agents still need human judgment, he argues, and he now focuses his reviews on architecture rather than line-by-line checks."

The workload also wears on him. Juggling six conversations about six features, he says, is more draining than coding ever was. He recalls rushing to answer agents' questions as if they were human colleagues, a habit he has since tempered. He also worries about the loss of craft, noting that many programmers valued the process of shaping code, not just the output.

Search and learning, but not product shopping

Generative AI is now his default for research and search. Google's AI summary is almost always the first or second result, he says, and it is usually decent. It is not perfect. In one day, it gave him incorrect URLs for NPR's RSS feeds. Still, he says traditional search often turned up nothing better, and AI helped him find a specific emoji, a technical term and a set of hardware specifications.

He has also come to use language models as teachers. He has had one walk him through fine-tuning a model to talk like a pirate, practice responding to an emergency medical call he heard on the radio, and explain a complex legal document. He has even had Claude quiz him on code an agent wrote so he could check his understanding.

Can AI tools teach as well as a human tutor?

His concern, though, is that these tools must be prompted to teach. Models are tuned to complete tasks for users, he says, and they often forget they are supposed to act as tutors. He also says accuracy has improved since last year, partly because he has learned which tasks the models handle well.

For product shopping, he remains unconvinced. He does not trust AI to recommend financial products or furniture, and he finds ads inside chatbot conversations unsettling. For literature searches, he now starts with a language model, but he says it can support almost any premise and may carry biases. Human experts remain more reliable. When he researched car touchscreens, he consulted a former colleague from a Georgia Tech lab rather than rely on the tools.

His research work has exposed the limits. While building a model to detect when a user has finished speaking, he says AI helped him find useful research and datasets, then led him astray when he asked about alternative architectures. He also spent days trying to get language models to help annotate training data, and a single discussion with a human teammate solved the problem.

Writing, art and the question of voice

Is AI good enough for writing and editing?

He still does not use AI to write from scratch or to outline his own work. He allows agents to document experiments and write readme files for code they generated, but he rewrites much of that prose himself. He says a few friends who asked AI to edit job applications received sycophantic, generic feedback, and that the tools tend to flatten strong writing. He finds them useful mainly for finding the right word.

He has stopped generating images. He says AI art is now everywhere, from advertisements to restaurant menus and store signs, and he calls its spread gross. He still enjoys some AI-generated music, especially songs that are not commercial. Video generation, in his view, has so far produced mostly misleading content, including history videos that recommended more low-quality clips.

He says he has drafted a personal AI policy, a document that sets out how he wants to use the tools. He is also working on an explicit policy for his website. He cancelled his OpenAI subscription about a month ago and now pays for Claude Pro at $17 a month and GitHub Copilot Pro at $10 a month, using Copilot solely for autocomplete. He says he expects the policy to change as the tools do.

Dieser Artikel wurde mit Unterstützung von KI verfasst.
News Factory APP - agentische News für besseres SEO & AEO.