From Vibe Coding to Agentic Engineering
Central idea
AI can produce code quickly, but engineers still own the understanding, judgment, and system-level decisions that make software dependable.
Summary
Software is entering a third era: natural language can now instruct a programmable model, not only a conventional computer. This dramatically lowers the cost of trying an idea. Yet capability is jagged—models can solve an impressive task and then miss an obvious constraint. Vibe coding raises the floor for creating software; agentic engineering raises the ceiling by pairing capable agents with explicit goals, verification, and human judgment. The engineer’s job moves upward without disappearing: specify the real problem, shape the architecture, inspect evidence, and decide when the result is trustworthy.
Key ideas
Verification is the dividing line
Agents improve fastest where outputs can be checked cheaply with tests, types, builds, simulations, or other objective feedback.
Delegation is not abdication
You may outsource implementation steps, but not your responsibility to understand the system and its risks.
Fundamentals gain leverage
Clear models of data, interfaces, failure modes, and trade-offs help you direct agents and catch plausible-looking mistakes.
A practical agentic loop
Specify the outcome and constraints.
→Let the agent propose and implement.
→Verify with executable evidence.
→Review the design and refine the brief.
Why it matters now
AI rewards engineers who can form precise intent and evaluate systems. Faster code generation makes weak judgment more expensive, not less important.
Continue with the original
This short version preserves the main argument. Follow the source for the complete talk, article, or book.