Seonglae ChoAI Coding Agent
The human role shifts from "coding" to environment design, specification writing, and feedback loop construction.
- Project Folder (Repository) = System of Record: All knowledge must exist explicitly within the repo.
AGENTS.mdshould be kept short, with actual knowledge distributed indocs/structure.- Documents are also mechanically verified with CI and lint.
- Agent Legibility First: Code style prioritizes structures that agents can easily understand over human preferences.
- Architecture layer structure: Types → Config → Repo → Service → Runtime → UI
- Dependency direction enforced by lint. Rules are enforced by code, not documentation.
- Agents directly handle UI, logs, and metrics
- Chrome DevTools Protocol integration.
- LogQL / PromQL access available, feedback loop must be automated
Problem for agent is Language Model Context Size
The advent of large language models could potentially reduce software development costs to nothing, sparking a rapid and diverse growth in software akin to the content boom or Cambrian explosion.
LLMs should be used in conjunction with other tools to prevent the human review process from becoming a bottleneck.
Existing coding benchmarks are good at measuring performance on tasks with clear correct answers, but they fail to capture what matters in real development: collaboration with people, productivity, judgment or taste, and the cost of revisions. Coding agents should be compared by metrics such as how many user interventions they require, how many tokens they consume, and how strong their UI judgment is.
AI Coding Agents
AI Coder Models
Code Quality leaderboard
Current limitations
- Stop Digging; Know Your Limits
- Mise en Place
- Scientific Debugging
- The tail wagging the dog
- Consistent formatting
- Read the Docs
- Use Static Types
Leaderboard
PR workflow integration Git workflow
Designing tools for developers means designing for LLMs too
Most large language models (LLMs) aren't great at using less popular frameworks.
Mac usage bar
usage stats
