We have officially entered into agentic engineering. The story has transformed into new kind of SDLC and I am in Go Gala since I have studied in 5 days Kaggle x Google Intensive vibe coding course that Software development is changed due to use of AI extensively. While AI do the heavy lifting writing codes and generating test cases, we need spec-driven development which indicates code is now disposable as AI can write it with just one prompt.However writing good specifications can make your software development errorless. If the requirements change you do not need to collapse infrastructure and rebuild it to exhaust your energy. Agents use wonderful skills, tools and LLM as a brain as part of Agentic AI.
A good specification can be hybrid markdown and YAML combined as discussed in the paper because they have 51.9% parsing accuracy. Developers do skip reasoning JSON format hence showcasing Agentic engineering optimal economics.
The other type is Behaviour driven development which includes natural language to structure human intention and ideas on a Gherkin syntax
"Scenario / Given / When / Then" template.
Always specify why are you doing than what during specification and break everything in piecewise which builds the agent reasoning and chances of your rapid development not getting stuck is minimised.
Different roles have different specifications and prompts either you are an architect, a builder or a forensic specialist; your specifications shall be distinct and bring you closer to the work being completed.
Even agents like Google antigravity provide the facility of terminal sandboxing where low-privileged containers can oversee the agent actions, even if the agent is acting dangerously everything outside of the radius will not be damaged.This can be helpful if you want to run edge cases for your software.
Embed autogenerated tests in the codebase will back fail-safe iteration and human in the loop comes into the play for even the higher stakes operation.
As in the paper AI users are facing approval fatigue when agent asks for literally each turn as compared to the non-AI users. For this we should having specs too.
Writing test specs is not a piece of cake but it does not reveal the full picture means it does show the errors but not the behavioural drift of the agent for example score the quality of the behaviour of the agent by LLM as a judge.It is known as evaluation of the agent which is highly needed if you want next level engineering and not just vibe coding.
There were numerous learnings from the course that began with harness engineering implying agent must have right skills and tools to execute the developer task and agent skills was a great read too.
Top comments (0)