Java applets from the 2000s no longer run in any modern browser. Yet thousands of web projects still rely on aging code that nobody has the time or budget to rewrite. This is exactly the problem AI coding agents are starting to solve, and an unexpected example has just demonstrated it in very concrete terms.
When a Math Genius Adopts AI Coding Agents
On July 11, 2026, Terence Tao, Fields Medal winner and one of the most recognized mathematicians in the world, published a post on his personal blog. In it, he describes how he used modern coding agents to migrate old Java applets into a maintainable repository. The starting observation is straightforward: these applets, which he had developed manually in another era, became unusable as web standards evolved.
This isn’t a professional developer speaking. This is someone for whom programming is a tool, not a career. And that’s precisely what makes this account interesting.
The post quickly gathered 422 points and 126 comments on Hacker News, well above average. Tao’s reputation certainly amplified the reach, but the enthusiasm also reflects genuine interest in practical use cases for these tools on legacy code.
What Legacy Code Migration Reveals About the State of Agents
Manually programming interactive web content was time-consuming, even for a mind as sharp as Tao’s. Coding agents change that equation. They let you describe an expected behavior and get working code back, without needing to master every syntactic detail of a modern framework.
For technical teams, the use case is even more direct:
- Rewriting Java applets or Flash scripts in modern JavaScript
- Converting Python 2 scripts to Python 3 with dependency management
- Migrating from jQuery to reactive frameworks without a full rewrite
- Automatically documenting code that has no existing comments
- Detecting security vulnerabilities in aging code before migration
In each of these cases, the coding agent acts as a translator between the old world and the new. The developer stays in control but delegates the repetitive and tedious parts.
The CLI Agent Ecosystem: A Useful Map
The GitHub repository bradAGI/awesome-cli-coding-agents provides a structured overview of coding agents available via command line. It distinguishes three main categories.
Open Source Tools
Aider, Goose, and OpenCode are among the go-to references. They integrate directly into the terminal, remain configurable, and keep data under local control. For sensitive projects or regulated environments, this is often the primary selection criterion.
Platform Agents
Claude Code (Anthropic), Codex (OpenAI), and Gemini CLI (Google) offer more advanced capabilities, backed by large-scale models. The user experience is often smoother, but the dependency on a third-party provider is real.
Orchestrators and Parallel Runners
Some tools allow running multiple agents in parallel or chaining complex tasks together. This is useful for large-scale migrations where several components need to be processed simultaneously.
This landscape evolves quickly. GitHub star counts give a rough sense of adoption, but they’re no substitute for testing on a real project.
The Data Sovereignty Question: A Quiet Warning Sign
A recent incident involving Claude Code deserves a clear mention. An issue opened on Anthropic’s official repository (anthropics/claude-code/issues/62476) reveals that the tool automatically deletes conversation transcripts after 30 days. Anthropic declined to fix this behavior.
For personal use, that’s an acceptable inconvenience. For professional or scientific use, it’s a different story. A coding agent transcript is a record of technical decisions. It’s audit history. In regulated contexts (healthcare, finance, defense), its automatic disappearance creates a real compliance problem.
This kind of incident illustrates a broader tension: proprietary agents offer the best capabilities but impose their own rules. Open source agents offer control but require more integration effort. There’s no universal answer, only a tradeoff to make based on context.
Key Takeaways
- AI coding agents make legacy code migration accessible to non-specialist profiles, as illustrated by Terence Tao’s experience with his Java applets.
- The ecosystem is mature and well mapped, spanning open source tools (Aider, Goose) and platform agents (Claude Code, Codex, Gemini CLI).
- Data sovereignty remains a significant friction point: the automatic deletion of Claude Code transcripts is a documented example.
- For sensitive migrations, favor open source agents or solutions hosted on infrastructure you control.
- Adoption by high-level scientific profiles validates the accessibility of these tools, but that doesn’t replace a rigorous technical evaluation before deployment.
If you’re working on a modernization project or are interested in coding agents, feel free to discuss it or follow the blog for upcoming analyses.
Sources
- Terence Tao’s post, July 11, 2026: a firsthand account of his migration process using AI agents
- bradAGI/awesome-cli-coding-agents: the reference repository listing active CLI agents
- Claude Code issue on transcript deletion: primary source on the Anthropic data retention incident