01 · Outcome
Start with the result that should improve.
The starting point is the user, the current workflow, and the useful result. That gives the agent a clear boundary before any model, framework, or provider is chosen.
The goal is focused automation inside a real product or operating process, not an isolated AI demonstration.
02 · System
Build the smallest complete system.
A useful agent usually needs more than a model call. I combine the product surface, prompting workflows, tools, APIs, data flows, integrations, and infrastructure required for a working first version.
- Connect models to the context and tools required by the workflow.
- Keep product and system boundaries explicit.
- Integrate with the software the team already uses.
- Leave a practical path to deploy, observe, and improve the system.
03 · Relevant proof
LLM-assisted review inside an existing delivery workflow.
One confirmed example is an LLM-powered code review assistant integrated into GitLab CI/CD. The work included structured prompting, tool and function calling, and a provider API layer using OpenAI and Anthropic models.
Read the code review case study