The spectrum of agentic coding
@ykdojo at @daftengine
Vibe coding and traditional software engineering are not opposites
Four levels of agentic coding
Level 1: Vibe coding
- Let AI go wild
- Forget the code even exists
- Error logs? Just throw them at the AI
- Version control? What's that?
Fine for one-off projects
Don't expect it to be maintainable
Level 2: Agentic coding with discipline
- Version control - Git and GitHub
- File-level understanding
- Basic security precautions
- Testing - make sure it actually works
Level 3: Agentic software engineering
- Tests you actually verify
- Pre-commit hooks
- CI jobs for continuous testing
- Function-level understanding
Additional quality control measures
- Careful manual testing
- End-to-end tests
- One-shot AI code reviews
- Regression checks
Level 4: High-quality software engineering
- Line-by-line understanding
- Self-reflection loops
- Interactive AI code reviews
Live terminal demo
To run locally: cd ~/Desktop/projects/librarian2 && ttyd -i 127.0.0.1 -p 7777 -W tmux new -A -s demo \; set status off \; set mouse on
github.com/Eventual-Inc/librarian2/pull/2
- Advanced research methods
- AI-powered quality control
- Accelerated learning
| Level |
Quality |
Speed |
Testing |
Code Review |
| Vibe Coding |
Low |
Lightning |
Bare minimum |
None |
| Agentic Coding with Discipline |
Okay |
Faster |
Sufficient |
Brief |
| Agentic Software Engineering |
Medium |
Fast |
Thorough |
One-shot |
| High-Quality Software Engineering |
High |
Decent |
Rock solid |
Interactive |
Master all four levels
Be flexible
Sometimes vibe coding is fine
Quick prototypes, one-off scripts, experiments
Sometimes you need high quality
Large-scale systems, mission-critical, medical software
The real problem with "slop"
more tokens == s**t?
more tokens != s**t
Live terminal demo
To run locally: cd ~/Desktop/projects/librarian2 && ttyd -i 127.0.0.1 -p 7777 -W tmux new -A -s demo \; set status off \; set mouse on
45 Claude Code tips
eventual.ai/agents
LinkedIn
