Bilingual

Wicked Features
5 days ago
- Wicked features are requirements that must be considered when building any other feature, dominating complexity in large tech companies.
- Examples of wicked features include adding new user types, on-premise SaaS versions, data sharding, data locality support, and internationalization (i18n).
- Wicked features are hard because they interact with each other like rules in the Password Game, requiring coordination and often leading to underestimation by engineers.
- Some wicked features are unavoidable due to high-paying enterprise demands (e.g., on-premise offerings, data locality), while others stem from overengineering or poor design choices.
- Good engineers should prevent unnecessary wicked features and limit their impact through sensible factoring to avoid polluting the entire system.
Anthropic runs large-scale code migrations with Claude Code
2 days ago
- Code migrations that used to take years can now be completed in weeks or even days using AI agents like Claude Code.
- Anthropic successfully migrated large codebases, including Bun's million-line Zig-to-Rust port and a Python-to-TypeScript migration of 165,000 lines.
- The core insight is to fix the process (loop) that produces the code, not the code itself.
- A six-step process is used: establish a judge, create rulebook and gap inventory, run a pilot, implement with multi-agent loops, compile and smoke test, and verify behavior parity.
- Key practices include using adversarial reviewers, refining rulebooks iteratively, and leveraging automated workflows to handle dependencies and errors systematically.
- Migrations now cost significantly less (e.g., $165,000 for Bun) but still require a business case, such as performance improvements or reducing technical debt.
- Successful migrations result in measurable benefits, like reduced memory usage, smaller binaries, and faster execution times.
Agent swarms and the new model economics
21 hours ago
- Agent swarms were developed to tackle large-scale tasks by dividing work into tree-like structures with planners and workers.
- New swarm improvements included a custom version control system handling high commit rates and mechanisms to address coordination failures like split-brain design and merge conflicts.
- Experiments showed the new swarm outperformed the old in tasks like building SQLite in Rust, achieving up to 85% test suite pass rates with better efficiency and lower costs.
- Cost analysis revealed significant savings by using frontier models for planning and cheaper models for execution, with hybrid configurations reducing worker costs dramatically.
- The swarm operates like a compiler for intent, translating high-level specs into executable work, with future potential to enhance software engineering through better intent description.

Hasty Briefsbeta

#software engineering

Wicked Features

Anthropic runs large-scale code migrations with Claude Code

Agent swarms and the new model economics